Hi
This patch series adds copy offload (the XCOPY command) to the block
layer, SCSI subsystems and device mapper.
The principle of operation is this:
We create two bios with REQ_COPY flag, one for read and one for write. The
bios have no data pages and they both point to the same bio_copy
structure. The bios may be submitted to the same logical volume or to
different logical volumes. The bios go independently through the device
mapper stack and when they both reach the physical device (the function
blk_queue_bio), the bio pair is collected and the XCOPY request is sent to
the SCSI device.
Mikulas
This is Martin Petersen's xcopy patch
(https://git.kernel.org/cgit/linux/kernel/git/mkp/linux.git/commit/?h=xcopy&id=0bdeed274e16b3038a851552188512071974eea8)
with some bug fixes, ported to the current kernel.
This patch makes it possible to use the SCSI XCOPY command.
We create a bio that has REQ_COPY flag in bi_rw and a bi_copy structure
that defines the source device. The target device is defined in the
bi_bdev and bi_iter.bi_sector.
There is a new BLKCOPY ioctl that makes it possible to use XCOPY from
userspace. The ioctl argument is a pointer to an array of four uint64_t
values.
The first value is a source byte offset, the second value is a destination
byte offset, the third value is byte length. The forth value is written by
the kernel and it represents the number of bytes that the kernel actually
copied.
Signed-off-by: Martin K. Petersen <[email protected]>
Signed-off-by: Mikulas Patocka <[email protected]>
---
Documentation/ABI/testing/sysfs-block | 9 +
block/bio.c | 2
block/blk-core.c | 5
block/blk-lib.c | 95 +++++++++++
block/blk-merge.c | 11 -
block/blk-settings.c | 13 +
block/blk-sysfs.c | 11 +
block/compat_ioctl.c | 1
block/ioctl.c | 50 ++++++
drivers/scsi/scsi.c | 57 +++++++
drivers/scsi/sd.c | 271 +++++++++++++++++++++++++++++++++-
drivers/scsi/sd.h | 4
include/linux/bio.h | 9 -
include/linux/blk_types.h | 14 +
include/linux/blkdev.h | 15 +
include/scsi/scsi_device.h | 3
include/uapi/linux/fs.h | 1
17 files changed, 557 insertions(+), 14 deletions(-)
Index: linux-4.4-rc4/Documentation/ABI/testing/sysfs-block
===================================================================
--- linux-4.4-rc4.orig/Documentation/ABI/testing/sysfs-block 2015-12-10 17:03:59.000000000 +0100
+++ linux-4.4-rc4/Documentation/ABI/testing/sysfs-block 2015-12-10 17:04:30.000000000 +0100
@@ -235,3 +235,12 @@ Description:
write_same_max_bytes is 0, write same is not supported
by the device.
+
+What: /sys/block/<disk>/queue/copy_max_bytes
+Date: January 2014
+Contact: Martin K. Petersen <[email protected]>
+Description:
+ Devices that support copy offloading will set this value
+ to indicate the maximum buffer size in bytes that can be
+ copied in one operation. If the copy_max_bytes is 0 the
+ device does not support copy offload.
Index: linux-4.4-rc4/block/blk-core.c
===================================================================
--- linux-4.4-rc4.orig/block/blk-core.c 2015-12-10 17:03:59.000000000 +0100
+++ linux-4.4-rc4/block/blk-core.c 2015-12-10 17:04:30.000000000 +0100
@@ -1957,6 +1957,11 @@ generic_make_request_checks(struct bio *
goto end_io;
}
+ if (bio->bi_rw & REQ_COPY && !bdev_copy_offload(bio->bi_bdev)) {
+ err = -EOPNOTSUPP;
+ goto end_io;
+ }
+
/*
* Various block parts want %current->io_context and lazy ioc
* allocation ends up trading a lot of pain for a small amount of
Index: linux-4.4-rc4/block/blk-lib.c
===================================================================
--- linux-4.4-rc4.orig/block/blk-lib.c 2015-12-10 17:03:59.000000000 +0100
+++ linux-4.4-rc4/block/blk-lib.c 2015-12-10 17:04:30.000000000 +0100
@@ -299,3 +299,98 @@ int blkdev_issue_zeroout(struct block_de
return __blkdev_issue_zeroout(bdev, sector, nr_sects, gfp_mask);
}
EXPORT_SYMBOL(blkdev_issue_zeroout);
+
+/**
+ * blkdev_issue_copy - queue a copy same operation
+ * @src_bdev: source blockdev
+ * @src_sector: source sector
+ * @dst_bdev: destination blockdev
+ * @dst_sector: destination sector
+ * @nr_sects: number of sectors to copy
+ * @gfp_mask: memory allocation flags (for bio_alloc)
+ *
+ * Description:
+ * Copy a block range from source device to target device.
+ */
+int blkdev_issue_copy(struct block_device *src_bdev, sector_t src_sector,
+ struct block_device *dst_bdev, sector_t dst_sector,
+ unsigned int nr_sects, gfp_t gfp_mask)
+{
+ DECLARE_COMPLETION_ONSTACK(wait);
+ struct request_queue *sq = bdev_get_queue(src_bdev);
+ struct request_queue *dq = bdev_get_queue(dst_bdev);
+ unsigned int max_copy_sectors;
+ struct bio_batch bb;
+ int ret = 0;
+
+ if (!sq || !dq)
+ return -ENXIO;
+
+ max_copy_sectors = min(sq->limits.max_copy_sectors,
+ dq->limits.max_copy_sectors);
+
+ if (max_copy_sectors == 0)
+ return -EOPNOTSUPP;
+
+ if (src_sector + nr_sects < src_sector ||
+ dst_sector + nr_sects < dst_sector)
+ return -EINVAL;
+
+ /* Do not support overlapping copies */
+ if (src_bdev == dst_bdev &&
+ abs((u64)dst_sector - (u64)src_sector) < nr_sects)
+ return -EOPNOTSUPP;
+
+ atomic_set(&bb.done, 1);
+ bb.error = 0;
+ bb.wait = &wait;
+
+ while (nr_sects) {
+ struct bio *bio;
+ struct bio_copy *bc;
+ unsigned int chunk;
+
+ bc = kmalloc(sizeof(struct bio_copy), gfp_mask);
+ if (!bc) {
+ ret = -ENOMEM;
+ break;
+ }
+
+ bio = bio_alloc(gfp_mask, 1);
+ if (!bio) {
+ kfree(bc);
+ ret = -ENOMEM;
+ break;
+ }
+
+ chunk = min(nr_sects, max_copy_sectors);
+
+ bio->bi_iter.bi_sector = dst_sector;
+ bio->bi_iter.bi_size = chunk << 9;
+ bio->bi_end_io = bio_batch_end_io;
+ bio->bi_bdev = dst_bdev;
+ bio->bi_private = &bb;
+ bio->bi_copy = bc;
+
+ bc->bic_bdev = src_bdev;
+ bc->bic_sector = src_sector;
+
+ atomic_inc(&bb.done);
+ submit_bio(REQ_WRITE | REQ_COPY, bio);
+
+ src_sector += chunk;
+ dst_sector += chunk;
+ nr_sects -= chunk;
+ }
+
+ /* Wait for bios in-flight */
+ if (!atomic_dec_and_test(&bb.done))
+ wait_for_completion_io(&wait);
+
+ if (likely(!ret))
+ ret = bb.error;
+
+ return ret;
+}
+EXPORT_SYMBOL(blkdev_issue_copy);
+
Index: linux-4.4-rc4/block/blk-merge.c
===================================================================
--- linux-4.4-rc4.orig/block/blk-merge.c 2015-12-10 17:03:59.000000000 +0100
+++ linux-4.4-rc4/block/blk-merge.c 2015-12-10 17:04:30.000000000 +0100
@@ -145,7 +145,9 @@ void blk_queue_split(struct request_queu
struct bio *split, *res;
unsigned nsegs;
- if ((*bio)->bi_rw & REQ_DISCARD)
+ if ((*bio)->bi_rw & REQ_COPY)
+ return;
+ else if ((*bio)->bi_rw & REQ_DISCARD)
split = blk_bio_discard_split(q, *bio, bs, &nsegs);
else if ((*bio)->bi_rw & REQ_WRITE_SAME)
split = blk_bio_write_same_split(q, *bio, bs, &nsegs);
@@ -185,10 +187,7 @@ static unsigned int __blk_recalc_rq_segm
* This should probably be returning 0, but blk_add_request_payload()
* (Christoph!!!!)
*/
- if (bio->bi_rw & REQ_DISCARD)
- return 1;
-
- if (bio->bi_rw & REQ_WRITE_SAME)
+ if (bio->bi_rw & (REQ_DISCARD | REQ_WRITE_SAME | REQ_COPY))
return 1;
fbio = bio;
@@ -361,7 +360,7 @@ static int __blk_bios_map_sg(struct requ
nsegs = 0;
cluster = blk_queue_cluster(q);
- if (bio->bi_rw & REQ_DISCARD) {
+ if (bio->bi_rw & (REQ_DISCARD | REQ_COPY)) {
/*
* This is a hack - drivers should be neither modifying the
* biovec, nor relying on bi_vcnt - but because of
Index: linux-4.4-rc4/block/blk-settings.c
===================================================================
--- linux-4.4-rc4.orig/block/blk-settings.c 2015-12-10 17:03:59.000000000 +0100
+++ linux-4.4-rc4/block/blk-settings.c 2015-12-10 17:04:30.000000000 +0100
@@ -95,6 +95,7 @@ void blk_set_default_limits(struct queue
BLK_SAFE_MAX_SECTORS;
lim->chunk_sectors = 0;
lim->max_write_same_sectors = 0;
+ lim->max_copy_sectors = 0;
lim->max_discard_sectors = 0;
lim->max_hw_discard_sectors = 0;
lim->discard_granularity = 0;
@@ -298,6 +299,18 @@ void blk_queue_max_write_same_sectors(st
EXPORT_SYMBOL(blk_queue_max_write_same_sectors);
/**
+ * blk_queue_max_copy_sectors - set max sectors for a single copy operation
+ * @q: the request queue for the device
+ * @max_copy_sectors: maximum number of sectors per copy operation
+ **/
+void blk_queue_max_copy_sectors(struct request_queue *q,
+ unsigned int max_copy_sectors)
+{
+ q->limits.max_copy_sectors = max_copy_sectors;
+}
+EXPORT_SYMBOL(blk_queue_max_copy_sectors);
+
+/**
* blk_queue_max_segments - set max hw segments for a request for this queue
* @q: the request queue for the device
* @max_segments: max number of segments
Index: linux-4.4-rc4/block/blk-sysfs.c
===================================================================
--- linux-4.4-rc4.orig/block/blk-sysfs.c 2015-12-10 17:04:01.000000000 +0100
+++ linux-4.4-rc4/block/blk-sysfs.c 2015-12-10 17:04:30.000000000 +0100
@@ -193,6 +193,11 @@ static ssize_t queue_write_same_max_show
(unsigned long long)q->limits.max_write_same_sectors << 9);
}
+static ssize_t queue_copy_max_show(struct request_queue *q, char *page)
+{
+ return sprintf(page, "%llu\n",
+ (unsigned long long)q->limits.max_copy_sectors << 9);
+}
static ssize_t
queue_max_sectors_store(struct request_queue *q, const char *page, size_t count)
@@ -443,6 +448,11 @@ static struct queue_sysfs_entry queue_wr
.show = queue_write_same_max_show,
};
+static struct queue_sysfs_entry queue_copy_max_entry = {
+ .attr = {.name = "copy_max_bytes", .mode = S_IRUGO },
+ .show = queue_copy_max_show,
+};
+
static struct queue_sysfs_entry queue_nonrot_entry = {
.attr = {.name = "rotational", .mode = S_IRUGO | S_IWUSR },
.show = queue_show_nonrot,
@@ -498,6 +508,7 @@ static struct attribute *default_attrs[]
&queue_discard_max_hw_entry.attr,
&queue_discard_zeroes_data_entry.attr,
&queue_write_same_max_entry.attr,
+ &queue_copy_max_entry.attr,
&queue_nonrot_entry.attr,
&queue_nomerges_entry.attr,
&queue_rq_affinity_entry.attr,
Index: linux-4.4-rc4/block/ioctl.c
===================================================================
--- linux-4.4-rc4.orig/block/ioctl.c 2015-12-10 17:03:59.000000000 +0100
+++ linux-4.4-rc4/block/ioctl.c 2015-12-10 17:04:30.000000000 +0100
@@ -249,6 +249,31 @@ static int blk_ioctl_zeroout(struct bloc
return blkdev_issue_zeroout(bdev, start, len, GFP_KERNEL, false);
}
+static int blk_ioctl_copy(struct block_device *bdev, uint64_t src_offset,
+ uint64_t dst_offset, uint64_t len)
+{
+ if (src_offset & 511)
+ return -EINVAL;
+ if (dst_offset & 511)
+ return -EINVAL;
+ if (len & 511)
+ return -EINVAL;
+ src_offset >>= 9;
+ dst_offset >>= 9;
+ len >>= 9;
+
+ if (unlikely(src_offset + len < src_offset) ||
+ unlikely(src_offset + len > (i_size_read(bdev->bd_inode) >> 9)))
+ return -EINVAL;
+
+ if (unlikely(dst_offset + len < dst_offset) ||
+ unlikely(dst_offset + len > (i_size_read(bdev->bd_inode) >> 9)))
+ return -EINVAL;
+
+ return blkdev_issue_copy(bdev, src_offset, bdev, dst_offset, len,
+ GFP_KERNEL);
+}
+
static int put_ushort(unsigned long arg, unsigned short val)
{
return put_user(val, (unsigned short __user *)arg);
@@ -513,6 +538,31 @@ int blkdev_ioctl(struct block_device *bd
BLKDEV_DISCARD_SECURE);
case BLKZEROOUT:
return blk_ioctl_zeroout(bdev, mode, arg);
+ case BLKCOPY: {
+ uint64_t range[4];
+ int ret;
+
+ range[3] = 0;
+
+ if (copy_to_user((void __user *)(arg + 24), &range[3], 8))
+ return -EFAULT;
+
+ if (!(mode & FMODE_WRITE))
+ return -EBADF;
+
+ if (copy_from_user(range, (void __user *)arg, 24))
+ return -EFAULT;
+
+ ret = blk_ioctl_copy(bdev, range[0], range[1], range[2]);
+ if (!ret) {
+ range[3] = range[2];
+ if (copy_to_user((void __user *)(arg + 24), &range[3], 8))
+ return -EFAULT;
+ }
+
+ return ret;
+ }
+
case HDIO_GETGEO:
return blkdev_getgeo(bdev, argp);
case BLKRAGET:
Index: linux-4.4-rc4/drivers/scsi/scsi.c
===================================================================
--- linux-4.4-rc4.orig/drivers/scsi/scsi.c 2015-12-10 17:04:00.000000000 +0100
+++ linux-4.4-rc4/drivers/scsi/scsi.c 2015-12-10 17:04:30.000000000 +0100
@@ -768,6 +768,62 @@ int scsi_get_vpd_page(struct scsi_device
EXPORT_SYMBOL_GPL(scsi_get_vpd_page);
/**
+ * scsi_lookup_naa - Lookup NAA descriptor in VPD page 0x83
+ * @sdev: The device to ask
+ *
+ * Copy offloading requires us to know the NAA descriptor for both
+ * source and target device. This descriptor is mandatory in the Device
+ * Identification VPD page. Locate this descriptor in the returned VPD
+ * data so we don't have to do lookups for every copy command.
+ */
+static void scsi_lookup_naa(struct scsi_device *sdev)
+{
+ unsigned char *buf = sdev->vpd_pg83;
+ unsigned int len = sdev->vpd_pg83_len;
+
+ if (buf[1] != 0x83 || get_unaligned_be16(&buf[2]) == 0) {
+ sdev_printk(KERN_ERR, sdev,
+ "%s: VPD page 0x83 contains no descriptors\n",
+ __func__);
+ return;
+ }
+
+ buf += 4;
+ len -= 4;
+
+ do {
+ unsigned int desig_len = buf[3] + 4;
+
+ /* Binary code set */
+ if ((buf[0] & 0xf) != 1)
+ goto skip;
+
+ /* Target association */
+ if ((buf[1] >> 4) & 0x3)
+ goto skip;
+
+ /* NAA designator */
+ if ((buf[1] & 0xf) != 0x3)
+ goto skip;
+
+ sdev->naa = buf;
+ sdev->naa_len = desig_len;
+
+ return;
+
+ skip:
+ buf += desig_len;
+ len -= desig_len;
+
+ } while (len > 0);
+
+ sdev_printk(KERN_ERR, sdev,
+ "%s: VPD page 0x83 NAA descriptor not found\n", __func__);
+
+ return;
+}
+
+/**
* scsi_attach_vpd - Attach Vital Product Data to a SCSI device structure
* @sdev: The device to ask
*
@@ -851,6 +907,7 @@ retry_pg83:
}
sdev->vpd_pg83_len = result;
sdev->vpd_pg83 = vpd_buf;
+ scsi_lookup_naa(sdev);
}
}
Index: linux-4.4-rc4/drivers/scsi/sd.c
===================================================================
--- linux-4.4-rc4.orig/drivers/scsi/sd.c 2015-12-10 17:04:00.000000000 +0100
+++ linux-4.4-rc4/drivers/scsi/sd.c 2015-12-10 17:04:30.000000000 +0100
@@ -101,6 +101,7 @@ MODULE_ALIAS_SCSI_DEVICE(TYPE_RBC);
static void sd_config_discard(struct scsi_disk *, unsigned int);
static void sd_config_write_same(struct scsi_disk *);
+static void sd_config_copy(struct scsi_disk *);
static int sd_revalidate_disk(struct gendisk *);
static void sd_unlock_native_capacity(struct gendisk *disk);
static int sd_probe(struct device *);
@@ -479,6 +480,48 @@ max_write_same_blocks_store(struct devic
}
static DEVICE_ATTR_RW(max_write_same_blocks);
+static ssize_t
+max_copy_blocks_show(struct device *dev, struct device_attribute *attr,
+ char *buf)
+{
+ struct scsi_disk *sdkp = to_scsi_disk(dev);
+
+ return snprintf(buf, 20, "%u\n", sdkp->max_copy_blocks);
+}
+
+static ssize_t
+max_copy_blocks_store(struct device *dev, struct device_attribute *attr,
+ const char *buf, size_t count)
+{
+ struct scsi_disk *sdkp = to_scsi_disk(dev);
+ struct scsi_device *sdp = sdkp->device;
+ unsigned long max;
+ int err;
+
+ if (!capable(CAP_SYS_ADMIN))
+ return -EACCES;
+
+ if (sdp->type != TYPE_DISK)
+ return -EINVAL;
+
+ err = kstrtoul(buf, 10, &max);
+
+ if (err)
+ return err;
+
+ if (max == 0)
+ sdp->no_copy = 1;
+ else if (max <= SD_MAX_COPY_BLOCKS) {
+ sdp->no_copy = 0;
+ sdkp->max_copy_blocks = max;
+ }
+
+ sd_config_copy(sdkp);
+
+ return count;
+}
+static DEVICE_ATTR_RW(max_copy_blocks);
+
static struct attribute *sd_disk_attrs[] = {
&dev_attr_cache_type.attr,
&dev_attr_FUA.attr,
@@ -490,6 +533,7 @@ static struct attribute *sd_disk_attrs[]
&dev_attr_thin_provisioning.attr,
&dev_attr_provisioning_mode.attr,
&dev_attr_max_write_same_blocks.attr,
+ &dev_attr_max_copy_blocks.attr,
&dev_attr_max_medium_access_timeouts.attr,
NULL,
};
@@ -879,6 +923,116 @@ static int sd_setup_write_same_cmnd(stru
return ret;
}
+static void sd_config_copy(struct scsi_disk *sdkp)
+{
+ struct request_queue *q = sdkp->disk->queue;
+ unsigned int logical_block_size = sdkp->device->sector_size;
+
+ if (sdkp->device->no_copy)
+ sdkp->max_copy_blocks = 0;
+
+ /* Segment descriptor 0x02 has a 64k block limit */
+ sdkp->max_copy_blocks = min(sdkp->max_copy_blocks,
+ (u32)SD_MAX_CSD2_BLOCKS);
+
+ blk_queue_max_copy_sectors(q, sdkp->max_copy_blocks *
+ (logical_block_size >> 9));
+}
+
+static int sd_setup_copy_cmnd(struct scsi_cmnd *cmd)
+{
+ struct request *rq = cmd->request;
+ struct scsi_device *src_sdp, *dst_sdp;
+ struct gendisk *src_disk;
+ struct request_queue *src_queue, *dst_queue;
+ sector_t src_lba, dst_lba;
+ unsigned int nr_blocks, buf_len, nr_bytes = blk_rq_bytes(rq);
+ int ret;
+ struct bio *bio = rq->bio;
+ struct page *page;
+ unsigned char *buf;
+
+ if (!bio->bi_copy)
+ return BLKPREP_KILL;
+
+ dst_sdp = scsi_disk(rq->rq_disk)->device;
+ dst_queue = rq->rq_disk->queue;
+ src_disk = bio->bi_copy->bic_bdev->bd_disk;
+ src_queue = src_disk->queue;
+ if (!src_queue ||
+ src_queue->make_request_fn != dst_queue->make_request_fn ||
+ src_queue->request_fn != dst_queue->request_fn ||
+ *(struct scsi_driver **)rq->rq_disk->private_data !=
+ *(struct scsi_driver **)src_disk->private_data)
+ return BLKPREP_KILL;
+ src_sdp = scsi_disk(src_disk)->device;
+
+ if (src_sdp->no_copy || dst_sdp->no_copy)
+ return BLKPREP_KILL;
+
+ if (src_sdp->sector_size != dst_sdp->sector_size)
+ return BLKPREP_KILL;
+
+ dst_lba = blk_rq_pos(rq) >> (ilog2(dst_sdp->sector_size) - 9);
+ src_lba = bio->bi_copy->bic_sector >> (ilog2(src_sdp->sector_size) - 9);
+ nr_blocks = blk_rq_sectors(rq) >> (ilog2(dst_sdp->sector_size) - 9);
+
+ page = alloc_page(GFP_ATOMIC | __GFP_ZERO);
+ if (!page)
+ return BLKPREP_DEFER;
+
+ buf = page_address(page);
+
+ /* Extended Copy (LID1) Parameter List (16 bytes) */
+ buf[0] = 0; /* LID */
+ buf[1] = 3 << 3; /* LID usage 11b */
+ put_unaligned_be16(32 + 32, &buf[2]); /* 32 bytes per E4 desc. */
+ put_unaligned_be32(28, &buf[8]); /* 28 bytes per B2B desc. */
+ buf += 16;
+
+ /* Source CSCD (32 bytes) */
+ buf[0] = 0xe4; /* Identification desc. */
+ memcpy(&buf[4], src_sdp->naa, src_sdp->naa_len);
+ buf += 32;
+
+ /* Destination CSCD (32 bytes) */
+ buf[0] = 0xe4; /* Identification desc. */
+ memcpy(&buf[4], dst_sdp->naa, dst_sdp->naa_len);
+ buf += 32;
+
+ /* Segment descriptor (28 bytes) */
+ buf[0] = 0x02; /* Block to block desc. */
+ put_unaligned_be16(0x18, &buf[2]); /* Descriptor length */
+ put_unaligned_be16(0, &buf[4]); /* Source is desc. 0 */
+ put_unaligned_be16(1, &buf[6]); /* Dest. is desc. 1 */
+ put_unaligned_be16(nr_blocks, &buf[10]);
+ put_unaligned_be64(src_lba, &buf[12]);
+ put_unaligned_be64(dst_lba, &buf[20]);
+
+ /* CDB */
+ cmd->cmd_len = 16;
+ memset(cmd->cmnd, 0, cmd->cmd_len);
+ cmd->cmnd[0] = EXTENDED_COPY;
+ cmd->cmnd[1] = 0; /* LID1 */
+ buf_len = 16 + 32 + 32 + 28;
+ put_unaligned_be32(buf_len, &cmd->cmnd[10]);
+ rq->timeout = SD_COPY_TIMEOUT;
+
+ rq->completion_data = page;
+ blk_add_request_payload(rq, page, buf_len);
+
+ cmd->transfersize = buf_len;
+ cmd->allowed = 0; /* don't retry */
+
+ rq->__data_len = buf_len;
+ ret = scsi_init_io(cmd);
+ rq->__data_len = nr_bytes;
+
+ if (ret != BLKPREP_OK)
+ __free_page(page);
+ return ret;
+}
+
static int sd_setup_flush_cmnd(struct scsi_cmnd *cmd)
{
struct request *rq = cmd->request;
@@ -1141,6 +1295,8 @@ static int sd_init_command(struct scsi_c
return sd_setup_discard_cmnd(cmd);
else if (rq->cmd_flags & REQ_WRITE_SAME)
return sd_setup_write_same_cmnd(cmd);
+ else if (rq->cmd_flags & REQ_COPY)
+ return sd_setup_copy_cmnd(cmd);
else if (rq->cmd_flags & REQ_FLUSH)
return sd_setup_flush_cmnd(cmd);
else
@@ -1151,7 +1307,7 @@ static void sd_uninit_command(struct scs
{
struct request *rq = SCpnt->request;
- if (rq->cmd_flags & REQ_DISCARD)
+ if (rq->cmd_flags & (REQ_DISCARD | REQ_COPY))
__free_page(rq->completion_data);
if (SCpnt->cmnd != rq->cmd) {
@@ -1768,7 +1924,8 @@ static int sd_done(struct scsi_cmnd *SCp
unsigned char op = SCpnt->cmnd[0];
unsigned char unmap = SCpnt->cmnd[1] & 8;
- if (req->cmd_flags & REQ_DISCARD || req->cmd_flags & REQ_WRITE_SAME) {
+ if (req->cmd_flags & REQ_DISCARD || req->cmd_flags & REQ_WRITE_SAME ||
+ req->cmd_flags & REQ_COPY) {
if (!result) {
good_bytes = blk_rq_bytes(req);
scsi_set_resid(SCpnt, 0);
@@ -1815,6 +1972,16 @@ static int sd_done(struct scsi_cmnd *SCp
/* INVALID COMMAND OPCODE or INVALID FIELD IN CDB */
if (sshdr.asc == 0x20 || sshdr.asc == 0x24) {
switch (op) {
+ case EXTENDED_COPY:
+ if ((SCpnt->cmnd[1] & 0x1f) == 0) {
+ sdkp->device->no_copy = 1;
+ sd_config_copy(sdkp);
+
+ good_bytes = 0;
+ req->__data_len = blk_rq_bytes(req);
+ req->cmd_flags |= REQ_QUIET;
+ }
+ break;
case UNMAP:
sd_config_discard(sdkp, SD_LBP_DISABLE);
break;
@@ -2797,6 +2964,105 @@ static void sd_read_write_same(struct sc
sdkp->ws10 = 1;
}
+static void sd_read_copy_operations(struct scsi_disk *sdkp,
+ unsigned char *buffer)
+{
+ struct scsi_device *sdev = sdkp->device;
+ struct scsi_sense_hdr sshdr;
+ unsigned char cdb[16];
+ unsigned int result, len, i;
+ bool b2b_desc = false, id_desc = false;
+
+ if (sdev->naa_len == 0)
+ return;
+
+ /* Verify that the device has 3PC set in INQUIRY response */
+ if (sdev->inquiry_len < 6 || (sdev->inquiry[5] & (1 << 3)) == 0)
+ return;
+
+ /* Receive Copy Operation Parameters */
+ memset(cdb, 0, 16);
+ cdb[0] = RECEIVE_COPY_RESULTS;
+ cdb[1] = 0x3;
+ put_unaligned_be32(SD_BUF_SIZE, &cdb[10]);
+
+ memset(buffer, 0, SD_BUF_SIZE);
+ result = scsi_execute_req(sdev, cdb, DMA_FROM_DEVICE,
+ buffer, SD_BUF_SIZE, &sshdr,
+ SD_TIMEOUT, SD_MAX_RETRIES, NULL);
+
+ if (!scsi_status_is_good(result)) {
+ sd_printk(KERN_ERR, sdkp,
+ "%s: Receive Copy Operating Parameters failed\n",
+ __func__);
+ return;
+ }
+
+ /* The RCOP response is a minimum of 44 bytes long. First 4
+ * bytes contain the length of the remaining buffer, i.e. 40+
+ * bytes. Trailing the defined fields is a list of supported
+ * descriptors. We need at least 2 descriptors to drive the
+ * target, hence 42.
+ */
+ len = get_unaligned_be32(&buffer[0]);
+ if (len < 42) {
+ sd_printk(KERN_ERR, sdkp, "%s: result too short (%u)\n",
+ __func__, len);
+ return;
+ }
+
+ if ((buffer[4] & 1) == 0) {
+ sd_printk(KERN_ERR, sdkp, "%s: does not support SNLID\n",
+ __func__);
+ return;
+ }
+
+ if (get_unaligned_be16(&buffer[8]) < 2) {
+ sd_printk(KERN_ERR, sdkp,
+ "%s: Need 2 or more CSCD descriptors\n", __func__);
+ return;
+ }
+
+ if (get_unaligned_be16(&buffer[10]) < 1) {
+ sd_printk(KERN_ERR, sdkp,
+ "%s: Need 1 or more segment descriptor\n", __func__);
+ return;
+ }
+
+ if (len - 40 != buffer[43]) {
+ sd_printk(KERN_ERR, sdkp,
+ "%s: Buffer len and descriptor count mismatch " \
+ "(%u vs. %u)\n", __func__, len - 40, buffer[43]);
+ return;
+ }
+
+ for (i = 44 ; i < len + 4 ; i++) {
+ if (buffer[i] == 0x02)
+ b2b_desc = true;
+
+ if (buffer[i] == 0xe4)
+ id_desc = true;
+ }
+
+ if (!b2b_desc) {
+ sd_printk(KERN_ERR, sdkp,
+ "%s: No block 2 block descriptor (0x02)\n",
+ __func__);
+ return;
+ }
+
+ if (!id_desc) {
+ sd_printk(KERN_ERR, sdkp,
+ "%s: No identification descriptor (0xE4)\n",
+ __func__);
+ return;
+ }
+
+ sdkp->max_copy_blocks = get_unaligned_be32(&buffer[16])
+ >> ilog2(sdev->sector_size);
+ sd_config_copy(sdkp);
+}
+
static int sd_try_extended_inquiry(struct scsi_device *sdp)
{
/* Attempt VPD inquiry if the device blacklist explicitly calls
@@ -2868,6 +3134,7 @@ static int sd_revalidate_disk(struct gen
sd_read_cache_type(sdkp, buffer);
sd_read_app_tag_own(sdkp, buffer);
sd_read_write_same(sdkp, buffer);
+ sd_read_copy_operations(sdkp, buffer);
}
sdkp->first_scan = 0;
Index: linux-4.4-rc4/drivers/scsi/sd.h
===================================================================
--- linux-4.4-rc4.orig/drivers/scsi/sd.h 2015-12-10 17:04:00.000000000 +0100
+++ linux-4.4-rc4/drivers/scsi/sd.h 2015-12-10 17:04:30.000000000 +0100
@@ -19,6 +19,7 @@
*/
#define SD_FLUSH_TIMEOUT_MULTIPLIER 2
#define SD_WRITE_SAME_TIMEOUT (120 * HZ)
+#define SD_COPY_TIMEOUT (120 * HZ)
/*
* Number of allowed retries
@@ -48,6 +49,8 @@ enum {
SD_MAX_XFER_BLOCKS = 0xffffffff,
SD_MAX_WS10_BLOCKS = 0xffff,
SD_MAX_WS16_BLOCKS = 0x7fffff,
+ SD_MAX_CSD2_BLOCKS = 0xffff,
+ SD_MAX_COPY_BLOCKS = 0xffffffff,
};
enum {
@@ -70,6 +73,7 @@ struct scsi_disk {
u32 opt_xfer_blocks;
u32 max_ws_blocks;
u32 max_unmap_blocks;
+ u32 max_copy_blocks;
u32 unmap_granularity;
u32 unmap_alignment;
u32 index;
Index: linux-4.4-rc4/include/linux/bio.h
===================================================================
--- linux-4.4-rc4.orig/include/linux/bio.h 2015-12-10 17:04:00.000000000 +0100
+++ linux-4.4-rc4/include/linux/bio.h 2015-12-10 17:04:30.000000000 +0100
@@ -106,7 +106,7 @@ static inline bool bio_has_data(struct b
{
if (bio &&
bio->bi_iter.bi_size &&
- !(bio->bi_rw & REQ_DISCARD))
+ !(bio->bi_rw & (REQ_DISCARD | REQ_COPY)))
return true;
return false;
@@ -249,8 +249,8 @@ static inline unsigned bio_segments(stru
struct bvec_iter iter;
/*
- * We special case discard/write same, because they interpret bi_size
- * differently:
+ * We special case discard/write same/copy, because they
+ * interpret bi_size differently:
*/
if (bio->bi_rw & REQ_DISCARD)
@@ -259,6 +259,9 @@ static inline unsigned bio_segments(stru
if (bio->bi_rw & REQ_WRITE_SAME)
return 1;
+ if (bio->bi_rw & REQ_COPY)
+ return 1;
+
bio_for_each_segment(bv, bio, iter)
segs++;
Index: linux-4.4-rc4/include/linux/blk_types.h
===================================================================
--- linux-4.4-rc4.orig/include/linux/blk_types.h 2015-12-10 17:04:00.000000000 +0100
+++ linux-4.4-rc4/include/linux/blk_types.h 2015-12-10 17:04:30.000000000 +0100
@@ -39,6 +39,11 @@ struct bvec_iter {
current bvec */
};
+struct bio_copy {
+ struct block_device *bic_bdev;
+ sector_t bic_sector;
+};
+
/*
* main unit of I/O for the block layer and lower layers (ie drivers and
* stacking drivers)
@@ -84,6 +89,7 @@ struct bio {
struct bio_integrity_payload *bi_integrity; /* data integrity */
#endif
};
+ struct bio_copy *bi_copy; /* TODO, use bi_integrity */
unsigned short bi_vcnt; /* how many bio_vec's */
@@ -156,6 +162,7 @@ enum rq_flag_bits {
__REQ_DISCARD, /* request to discard sectors */
__REQ_SECURE, /* secure discard (used with __REQ_DISCARD) */
__REQ_WRITE_SAME, /* write same block many times */
+ __REQ_COPY, /* copy block range */
__REQ_NOIDLE, /* don't anticipate more IO after this one */
__REQ_INTEGRITY, /* I/O includes block integrity payload */
@@ -201,6 +208,7 @@ enum rq_flag_bits {
#define REQ_PRIO (1ULL << __REQ_PRIO)
#define REQ_DISCARD (1ULL << __REQ_DISCARD)
#define REQ_WRITE_SAME (1ULL << __REQ_WRITE_SAME)
+#define REQ_COPY (1ULL << __REQ_COPY)
#define REQ_NOIDLE (1ULL << __REQ_NOIDLE)
#define REQ_INTEGRITY (1ULL << __REQ_INTEGRITY)
@@ -209,14 +217,14 @@ enum rq_flag_bits {
#define REQ_COMMON_MASK \
(REQ_WRITE | REQ_FAILFAST_MASK | REQ_SYNC | REQ_META | REQ_PRIO | \
REQ_DISCARD | REQ_WRITE_SAME | REQ_NOIDLE | REQ_FLUSH | REQ_FUA | \
- REQ_SECURE | REQ_INTEGRITY)
+ REQ_SECURE | REQ_INTEGRITY | REQ_COPY)
#define REQ_CLONE_MASK REQ_COMMON_MASK
-#define BIO_NO_ADVANCE_ITER_MASK (REQ_DISCARD|REQ_WRITE_SAME)
+#define BIO_NO_ADVANCE_ITER_MASK (REQ_DISCARD|REQ_WRITE_SAME|REQ_COPY)
/* This mask is used for both bio and request merge checking */
#define REQ_NOMERGE_FLAGS \
- (REQ_NOMERGE | REQ_STARTED | REQ_SOFTBARRIER | REQ_FLUSH | REQ_FUA | REQ_FLUSH_SEQ)
+ (REQ_NOMERGE | REQ_STARTED | REQ_SOFTBARRIER | REQ_FLUSH | REQ_FUA | REQ_FLUSH_SEQ | REQ_COPY)
#define REQ_RAHEAD (1ULL << __REQ_RAHEAD)
#define REQ_THROTTLED (1ULL << __REQ_THROTTLED)
Index: linux-4.4-rc4/include/linux/blkdev.h
===================================================================
--- linux-4.4-rc4.orig/include/linux/blkdev.h 2015-12-10 17:04:01.000000000 +0100
+++ linux-4.4-rc4/include/linux/blkdev.h 2015-12-10 17:04:30.000000000 +0100
@@ -265,6 +265,7 @@ struct queue_limits {
unsigned int max_discard_sectors;
unsigned int max_hw_discard_sectors;
unsigned int max_write_same_sectors;
+ unsigned int max_copy_sectors;
unsigned int discard_granularity;
unsigned int discard_alignment;
@@ -968,6 +969,8 @@ extern void blk_queue_max_discard_sector
unsigned int max_discard_sectors);
extern void blk_queue_max_write_same_sectors(struct request_queue *q,
unsigned int max_write_same_sectors);
+extern void blk_queue_max_copy_sectors(struct request_queue *q,
+ unsigned int max_copy_sectors);
extern void blk_queue_logical_block_size(struct request_queue *, unsigned short);
extern void blk_queue_physical_block_size(struct request_queue *, unsigned int);
extern void blk_queue_alignment_offset(struct request_queue *q,
@@ -1137,6 +1140,8 @@ extern int blkdev_issue_discard(struct b
sector_t nr_sects, gfp_t gfp_mask, unsigned long flags);
extern int blkdev_issue_write_same(struct block_device *bdev, sector_t sector,
sector_t nr_sects, gfp_t gfp_mask, struct page *page);
+extern int blkdev_issue_copy(struct block_device *, sector_t,
+ struct block_device *, sector_t, unsigned int, gfp_t);
extern int blkdev_issue_zeroout(struct block_device *bdev, sector_t sector,
sector_t nr_sects, gfp_t gfp_mask, bool discard);
static inline int sb_issue_discard(struct super_block *sb, sector_t block,
@@ -1340,6 +1345,16 @@ static inline unsigned int bdev_write_sa
return 0;
}
+static inline unsigned int bdev_copy_offload(struct block_device *bdev)
+{
+ struct request_queue *q = bdev_get_queue(bdev);
+
+ if (q)
+ return q->limits.max_copy_sectors;
+
+ return 0;
+}
+
static inline int queue_dma_alignment(struct request_queue *q)
{
return q ? q->dma_alignment : 511;
Index: linux-4.4-rc4/include/scsi/scsi_device.h
===================================================================
--- linux-4.4-rc4.orig/include/scsi/scsi_device.h 2015-12-10 17:04:00.000000000 +0100
+++ linux-4.4-rc4/include/scsi/scsi_device.h 2015-12-10 17:04:30.000000000 +0100
@@ -120,6 +120,8 @@ struct scsi_device {
unsigned char *vpd_pg83;
int vpd_pg80_len;
unsigned char *vpd_pg80;
+ unsigned char naa_len;
+ unsigned char *naa;
unsigned char current_tag; /* current tag */
struct scsi_target *sdev_target; /* used only for single_lun */
@@ -150,6 +152,7 @@ struct scsi_device {
unsigned use_10_for_ms:1; /* first try 10-byte mode sense/select */
unsigned no_report_opcodes:1; /* no REPORT SUPPORTED OPERATION CODES */
unsigned no_write_same:1; /* no WRITE SAME command */
+ unsigned no_copy:1; /* no copy offload */
unsigned use_16_for_rw:1; /* Use read/write(16) over read/write(10) */
unsigned skip_ms_page_8:1; /* do not use MODE SENSE page 0x08 */
unsigned skip_ms_page_3f:1; /* do not use MODE SENSE page 0x3f */
Index: linux-4.4-rc4/include/uapi/linux/fs.h
===================================================================
--- linux-4.4-rc4.orig/include/uapi/linux/fs.h 2015-12-10 17:04:00.000000000 +0100
+++ linux-4.4-rc4/include/uapi/linux/fs.h 2015-12-10 17:04:30.000000000 +0100
@@ -152,6 +152,7 @@ struct inodes_stat_t {
#define BLKSECDISCARD _IO(0x12,125)
#define BLKROTATIONAL _IO(0x12,126)
#define BLKZEROOUT _IO(0x12,127)
+#define BLKCOPY _IO(0x12,128)
#define BMAP_IOCTL 1 /* obsolete - kept for compatibility */
#define FIBMAP _IO(0x00,1) /* bmap access */
Index: linux-4.4-rc4/block/compat_ioctl.c
===================================================================
--- linux-4.4-rc4.orig/block/compat_ioctl.c 2015-12-10 17:04:00.000000000 +0100
+++ linux-4.4-rc4/block/compat_ioctl.c 2015-12-10 17:04:30.000000000 +0100
@@ -697,6 +697,7 @@ long compat_blkdev_ioctl(struct file *fi
* but we call blkdev_ioctl, which gets the lock for us
*/
case BLKRRPART:
+ case BLKCOPY:
return blkdev_ioctl(bdev, mode, cmd,
(unsigned long)compat_ptr(arg));
case BLKBSZSET_32:
Index: linux-4.4-rc4/block/bio.c
===================================================================
--- linux-4.4-rc4.orig/block/bio.c 2015-12-10 17:03:59.000000000 +0100
+++ linux-4.4-rc4/block/bio.c 2015-12-10 17:04:30.000000000 +0100
@@ -238,6 +238,8 @@ static void __bio_free(struct bio *bio)
{
bio_disassociate_task(bio);
+ kfree(bio->bi_copy);
+
if (bio_integrity(bio))
bio_integrity_free(bio);
}
This patch changes the architecture of xcopy so that two bios are used.
There used to be just one bio that held pointers to both source and
destination block device. However a bio with two block devices cannot
really be passed though block midlayer drivers (dm and md).
When we need to send the XCOPY command, we call the function
blkdev_issue_copy. This function creates two bios, the first with bi_rw
READ | REQ_COPY and the second WRITE | REQ_COPY. The bios have a pointer
to a common bi_copy structure.
These bios travel independently through the block device stack. When both
the bios reach the physical disk driver (the function blk_queue_bio), they
are paired, a request is made and the request is sent to the SCSI disk
driver.
It is possible that one of the bios reaches a device that doesn't support
XCOPY, in that case both bios are aborted with an error.
Note that there is no guarantee that the XCOPY command will succeed. If it
doesn't succeed, the caller is supposed to perform the copy manually.
Signed-off-by: Mikulas Patocka <[email protected]>
---
block/bio.c | 28 +++++++++++++++-
block/blk-core.c | 41 ++++++++++++++++++++++++
block/blk-lib.c | 78 +++++++++++++++++++++++++++++++++++++---------
drivers/scsi/sd.c | 7 +---
include/linux/blk_types.h | 12 +++++--
5 files changed, 142 insertions(+), 24 deletions(-)
Index: linux-4.4-rc4/block/blk-lib.c
===================================================================
--- linux-4.4-rc4.orig/block/blk-lib.c 2015-12-10 17:08:11.000000000 +0100
+++ linux-4.4-rc4/block/blk-lib.c 2015-12-10 17:59:40.000000000 +0100
@@ -300,6 +300,38 @@ int blkdev_issue_zeroout(struct block_de
}
EXPORT_SYMBOL(blkdev_issue_zeroout);
+static void bio_copy_end_io(struct bio *bio)
+{
+ struct bio_copy *bc = bio->bi_copy;
+ if (unlikely(bio->bi_error)) {
+ unsigned long flags;
+ int dir;
+ struct bio *other;
+
+ /* if the other bio is waiting for the pair, release it */
+ spin_lock_irqsave(&bc->spinlock, flags);
+ if (bc->error >= 0)
+ bc->error = bio->bi_error;
+ dir = bio_data_dir(bio);
+ other = bc->pair[dir ^ 1];
+ bc->pair[dir ^ 1] = NULL;
+ spin_unlock_irqrestore(&bc->spinlock, flags);
+ if (other) {
+ other->bi_error = bio->bi_error;
+ bio_endio(other);
+ }
+ }
+ bio_put(bio);
+ if (atomic_dec_and_test(&bc->in_flight)) {
+ struct bio_batch *bb = bc->private;
+ if (unlikely(bc->error < 0) && !ACCESS_ONCE(bb->error))
+ ACCESS_ONCE(bb->error) = bc->error;
+ kfree(bc);
+ if (atomic_dec_and_test(&bb->done))
+ complete(bb->wait);
+ }
+}
+
/**
* blkdev_issue_copy - queue a copy same operation
* @src_bdev: source blockdev
@@ -346,9 +378,9 @@ int blkdev_issue_copy(struct block_devic
bb.wait = &wait;
while (nr_sects) {
- struct bio *bio;
+ struct bio *read_bio, *write_bio;
struct bio_copy *bc;
- unsigned int chunk;
+ unsigned int chunk = min(nr_sects, max_copy_sectors);
bc = kmalloc(sizeof(struct bio_copy), gfp_mask);
if (!bc) {
@@ -356,27 +388,43 @@ int blkdev_issue_copy(struct block_devic
break;
}
- bio = bio_alloc(gfp_mask, 1);
- if (!bio) {
+ read_bio = bio_alloc(gfp_mask, 1);
+ if (!read_bio) {
kfree(bc);
ret = -ENOMEM;
break;
}
- chunk = min(nr_sects, max_copy_sectors);
-
- bio->bi_iter.bi_sector = dst_sector;
- bio->bi_iter.bi_size = chunk << 9;
- bio->bi_end_io = bio_batch_end_io;
- bio->bi_bdev = dst_bdev;
- bio->bi_private = &bb;
- bio->bi_copy = bc;
+ write_bio = bio_alloc(gfp_mask, 1);
+ if (!write_bio) {
+ bio_put(read_bio);
+ kfree(bc);
+ ret = -ENOMEM;
+ break;
+ }
- bc->bic_bdev = src_bdev;
- bc->bic_sector = src_sector;
+ atomic_set(&bc->in_flight, 2);
+ bc->error = 1;
+ bc->pair[0] = NULL;
+ bc->pair[1] = NULL;
+ bc->private = &bb;
+ spin_lock_init(&bc->spinlock);
+
+ read_bio->bi_iter.bi_sector = src_sector;
+ read_bio->bi_iter.bi_size = chunk << 9;
+ read_bio->bi_end_io = bio_copy_end_io;
+ read_bio->bi_bdev = src_bdev;
+ read_bio->bi_copy = bc;
+
+ write_bio->bi_iter.bi_sector = dst_sector;
+ write_bio->bi_iter.bi_size = chunk << 9;
+ write_bio->bi_end_io = bio_copy_end_io;
+ write_bio->bi_bdev = dst_bdev;
+ write_bio->bi_copy = bc;
atomic_inc(&bb.done);
- submit_bio(REQ_WRITE | REQ_COPY, bio);
+ submit_bio(READ | REQ_COPY, read_bio);
+ submit_bio(WRITE | REQ_COPY, write_bio);
src_sector += chunk;
dst_sector += chunk;
Index: linux-4.4-rc4/include/linux/blk_types.h
===================================================================
--- linux-4.4-rc4.orig/include/linux/blk_types.h 2015-12-10 17:08:11.000000000 +0100
+++ linux-4.4-rc4/include/linux/blk_types.h 2015-12-10 17:59:41.000000000 +0100
@@ -40,8 +40,16 @@ struct bvec_iter {
};
struct bio_copy {
- struct block_device *bic_bdev;
- sector_t bic_sector;
+ /*
+ * error == 1 - bios are waiting to be paired
+ * error == 0 - pair was issued
+ * error < 0 - error
+ */
+ int error;
+ atomic_t in_flight;
+ struct bio *pair[2];
+ void *private;
+ spinlock_t spinlock;
};
/*
Index: linux-4.4-rc4/block/bio.c
===================================================================
--- linux-4.4-rc4.orig/block/bio.c 2015-12-10 17:08:11.000000000 +0100
+++ linux-4.4-rc4/block/bio.c 2015-12-10 17:08:11.000000000 +0100
@@ -238,8 +238,6 @@ static void __bio_free(struct bio *bio)
{
bio_disassociate_task(bio);
- kfree(bio->bi_copy);
-
if (bio_integrity(bio))
bio_integrity_free(bio);
}
@@ -566,6 +564,7 @@ void __bio_clone_fast(struct bio *bio, s
bio_set_flag(bio, BIO_CLONED);
bio->bi_rw = bio_src->bi_rw;
bio->bi_iter = bio_src->bi_iter;
+ bio->bi_copy = bio_src->bi_copy;
bio->bi_io_vec = bio_src->bi_io_vec;
}
EXPORT_SYMBOL(__bio_clone_fast);
@@ -1711,6 +1710,28 @@ static inline bool bio_remaining_done(st
return false;
}
+static noinline_for_stack void bio_endio_copy(struct bio *bio)
+{
+ struct bio_copy *bc = bio->bi_copy;
+ struct bio *other = NULL;
+ unsigned long flags;
+ int dir;
+
+ spin_lock_irqsave(&bc->spinlock, flags);
+ dir = bio_data_dir(bio);
+ if (bc->pair[dir]) {
+ BUG_ON(bc->pair[dir] != bio);
+ other = bc->pair[dir ^ 1];
+ bc->pair[0] = bc->pair[1] = NULL;
+ }
+ spin_unlock_irqrestore(&bc->spinlock, flags);
+
+ if (other) {
+ other->bi_error = bio->bi_error;
+ bio_endio(other);
+ }
+}
+
/**
* bio_endio - end I/O on a bio
* @bio: bio
@@ -1726,6 +1747,9 @@ void bio_endio(struct bio *bio)
if (unlikely(!bio_remaining_done(bio)))
break;
+ if (unlikely((bio->bi_rw & REQ_COPY) != 0))
+ bio_endio_copy(bio);
+
/*
* Need to have a real endio function for chained bios,
* otherwise various corner cases will break (like stacking
Index: linux-4.4-rc4/block/blk-core.c
===================================================================
--- linux-4.4-rc4.orig/block/blk-core.c 2015-12-10 17:08:11.000000000 +0100
+++ linux-4.4-rc4/block/blk-core.c 2015-12-10 18:03:47.000000000 +0100
@@ -1681,6 +1681,33 @@ void init_request_from_bio(struct reques
blk_rq_bio_prep(req->q, req, bio);
}
+static noinline_for_stack struct bio *blk_queue_copy(struct bio *bio)
+{
+ struct bio_copy *bc = bio->bi_copy;
+ int dir, error;
+ struct bio *ret;
+
+ spin_lock_irq(&bc->spinlock);
+ error = bc->error;
+ if (unlikely(error < 0)) {
+ spin_unlock_irq(&bc->spinlock);
+ bio->bi_error = error;
+ bio_endio(bio);
+ return NULL;
+ }
+ dir = bio_data_dir(bio);
+ bc->pair[dir] = bio;
+ if (bc->pair[dir ^ 1]) {
+ ret = bc->pair[1];
+ bc->error = 0;
+ } else {
+ ret = NULL;
+ }
+ spin_unlock_irq(&bc->spinlock);
+
+ return ret;
+}
+
static blk_qc_t blk_queue_bio(struct request_queue *q, struct bio *bio)
{
const bool sync = !!(bio->bi_rw & REQ_SYNC);
@@ -1740,6 +1767,20 @@ static blk_qc_t blk_queue_bio(struct req
}
get_rq:
+ if (unlikely((bio->bi_rw & REQ_COPY) != 0)) {
+ if (unlikely(bio_sectors(bio) > q->limits.max_copy_sectors)) {
+ spin_unlock_irq(q->queue_lock);
+ bio->bi_error = -EOPNOTSUPP;
+ bio_endio(bio);
+ return BLK_QC_T_NONE;
+ }
+ spin_unlock_irq(q->queue_lock);
+ bio = blk_queue_copy(bio);
+ if (!bio)
+ return BLK_QC_T_NONE;
+ spin_lock_irq(q->queue_lock);
+ }
+
/*
* This sync check and mask will be re-done in init_request_from_bio(),
* but we need to set it earlier to expose the sync flag to the
Index: linux-4.4-rc4/drivers/scsi/sd.c
===================================================================
--- linux-4.4-rc4.orig/drivers/scsi/sd.c 2015-12-10 17:08:11.000000000 +0100
+++ linux-4.4-rc4/drivers/scsi/sd.c 2015-12-10 17:59:38.000000000 +0100
@@ -952,12 +952,9 @@ static int sd_setup_copy_cmnd(struct scs
struct page *page;
unsigned char *buf;
- if (!bio->bi_copy)
- return BLKPREP_KILL;
-
dst_sdp = scsi_disk(rq->rq_disk)->device;
dst_queue = rq->rq_disk->queue;
- src_disk = bio->bi_copy->bic_bdev->bd_disk;
+ src_disk = bio->bi_copy->pair[0]->bi_bdev->bd_disk;
src_queue = src_disk->queue;
if (!src_queue ||
src_queue->make_request_fn != dst_queue->make_request_fn ||
@@ -974,7 +971,7 @@ static int sd_setup_copy_cmnd(struct scs
return BLKPREP_KILL;
dst_lba = blk_rq_pos(rq) >> (ilog2(dst_sdp->sector_size) - 9);
- src_lba = bio->bi_copy->bic_sector >> (ilog2(src_sdp->sector_size) - 9);
+ src_lba = bio->bi_copy->pair[0]->bi_iter.bi_sector >> (ilog2(src_sdp->sector_size) - 9);
nr_blocks = blk_rq_sectors(rq) >> (ilog2(dst_sdp->sector_size) - 9);
page = alloc_page(GFP_ATOMIC | __GFP_ZERO);
This patch changes blkdev_issue_copy so that it returns the number of
copied sectors in the variable "copied".
The kernel makes best effort to copy as much data as possible, but because
of device mapper mapping, it may be possible that copying fails at some
stage. If we just returned the error number, the caller wouldn't know if
all or part of the operation failed and the caller would be required to
redo the whole copy operation.
We return the number of copied sectors so that the caller can skip these
sectors when doing the copy manually. On success (zero return code), the
number of copied sectors is equal to the number of requested sectors. On
error (negative return code), the number of copied sectors is smaller than
the number of requested sectors.
The number of copied bytes is returned as a fourth uint64_t argument in
the BLKCOPY ioctl.
Signed-off-by: Mikulas Patocka <[email protected]>
---
block/blk-lib.c | 30 +++++++++++++++++++++++++-----
block/ioctl.c | 25 ++++++++++++++++---------
include/linux/blk_types.h | 2 ++
include/linux/blkdev.h | 3 ++-
4 files changed, 45 insertions(+), 15 deletions(-)
Index: linux-4.4-rc4/block/blk-lib.c
===================================================================
--- linux-4.4-rc4.orig/block/blk-lib.c 2015-12-10 17:04:38.000000000 +0100
+++ linux-4.4-rc4/block/blk-lib.c 2015-12-10 17:04:40.000000000 +0100
@@ -324,8 +324,17 @@ static void bio_copy_end_io(struct bio *
bio_put(bio);
if (atomic_dec_and_test(&bc->in_flight)) {
struct bio_batch *bb = bc->private;
- if (unlikely(bc->error < 0) && !ACCESS_ONCE(bb->error))
- ACCESS_ONCE(bb->error) = bc->error;
+ if (unlikely(bc->error < 0)) {
+ u64 first_error;
+ if (!ACCESS_ONCE(bb->error))
+ ACCESS_ONCE(bb->error) = bc->error;
+ do {
+ first_error = atomic64_read(bc->first_error);
+ if (bc->offset >= first_error)
+ break;
+ } while (unlikely(atomic64_cmpxchg(bc->first_error,
+ first_error, bc->offset) != first_error));
+ }
kfree(bc);
if (atomic_dec_and_test(&bb->done))
complete(bb->wait);
@@ -346,7 +355,7 @@ static void bio_copy_end_io(struct bio *
*/
int blkdev_issue_copy(struct block_device *src_bdev, sector_t src_sector,
struct block_device *dst_bdev, sector_t dst_sector,
- unsigned int nr_sects, gfp_t gfp_mask)
+ sector_t nr_sects, gfp_t gfp_mask, sector_t *copied)
{
DECLARE_COMPLETION_ONSTACK(wait);
struct request_queue *sq = bdev_get_queue(src_bdev);
@@ -354,6 +363,11 @@ int blkdev_issue_copy(struct block_devic
unsigned int max_copy_sectors;
struct bio_batch bb;
int ret = 0;
+ atomic64_t first_error = ATOMIC64_INIT(nr_sects);
+ sector_t offset = 0;
+
+ if (copied)
+ *copied = 0;
if (!sq || !dq)
return -ENXIO;
@@ -377,10 +391,10 @@ int blkdev_issue_copy(struct block_devic
bb.error = 0;
bb.wait = &wait;
- while (nr_sects) {
+ while (nr_sects && !ACCESS_ONCE(bb.error)) {
struct bio *read_bio, *write_bio;
struct bio_copy *bc;
- unsigned int chunk = min(nr_sects, max_copy_sectors);
+ unsigned chunk = (unsigned)min(nr_sects, (sector_t)max_copy_sectors);
bc = kmalloc(sizeof(struct bio_copy), gfp_mask);
if (!bc) {
@@ -408,6 +422,8 @@ int blkdev_issue_copy(struct block_devic
bc->pair[0] = NULL;
bc->pair[1] = NULL;
bc->private = &bb;
+ bc->first_error = &first_error;
+ bc->offset = offset;
spin_lock_init(&bc->spinlock);
read_bio->bi_iter.bi_sector = src_sector;
@@ -429,12 +445,16 @@ int blkdev_issue_copy(struct block_devic
src_sector += chunk;
dst_sector += chunk;
nr_sects -= chunk;
+ offset += chunk;
}
/* Wait for bios in-flight */
if (!atomic_dec_and_test(&bb.done))
wait_for_completion_io(&wait);
+ if (copied)
+ *copied = min((sector_t)atomic64_read(&first_error), offset);
+
if (likely(!ret))
ret = bb.error;
Index: linux-4.4-rc4/include/linux/blk_types.h
===================================================================
--- linux-4.4-rc4.orig/include/linux/blk_types.h 2015-12-10 17:04:38.000000000 +0100
+++ linux-4.4-rc4/include/linux/blk_types.h 2015-12-10 17:04:40.000000000 +0100
@@ -49,6 +49,8 @@ struct bio_copy {
atomic_t in_flight;
struct bio *pair[2];
void *private;
+ atomic64_t *first_error;
+ sector_t offset;
spinlock_t spinlock;
};
Index: linux-4.4-rc4/include/linux/blkdev.h
===================================================================
--- linux-4.4-rc4.orig/include/linux/blkdev.h 2015-12-10 17:04:30.000000000 +0100
+++ linux-4.4-rc4/include/linux/blkdev.h 2015-12-10 17:04:40.000000000 +0100
@@ -1141,7 +1141,8 @@ extern int blkdev_issue_discard(struct b
extern int blkdev_issue_write_same(struct block_device *bdev, sector_t sector,
sector_t nr_sects, gfp_t gfp_mask, struct page *page);
extern int blkdev_issue_copy(struct block_device *, sector_t,
- struct block_device *, sector_t, unsigned int, gfp_t);
+ struct block_device *, sector_t, sector_t, gfp_t,
+ sector_t *);
extern int blkdev_issue_zeroout(struct block_device *bdev, sector_t sector,
sector_t nr_sects, gfp_t gfp_mask, bool discard);
static inline int sb_issue_discard(struct super_block *sb, sector_t block,
Index: linux-4.4-rc4/block/ioctl.c
===================================================================
--- linux-4.4-rc4.orig/block/ioctl.c 2015-12-10 17:04:30.000000000 +0100
+++ linux-4.4-rc4/block/ioctl.c 2015-12-10 17:04:40.000000000 +0100
@@ -250,8 +250,13 @@ static int blk_ioctl_zeroout(struct bloc
}
static int blk_ioctl_copy(struct block_device *bdev, uint64_t src_offset,
- uint64_t dst_offset, uint64_t len)
+ uint64_t dst_offset, uint64_t len, uint64_t *copied)
{
+ int ret;
+ sector_t copied_sec;
+
+ *copied = 0;
+
if (src_offset & 511)
return -EINVAL;
if (dst_offset & 511)
@@ -270,8 +275,12 @@ static int blk_ioctl_copy(struct block_d
unlikely(dst_offset + len > (i_size_read(bdev->bd_inode) >> 9)))
return -EINVAL;
- return blkdev_issue_copy(bdev, src_offset, bdev, dst_offset, len,
- GFP_KERNEL);
+ ret = blkdev_issue_copy(bdev, src_offset, bdev, dst_offset, len,
+ GFP_KERNEL, &copied_sec);
+
+ *copied = (uint64_t)copied_sec << 9;
+
+ return ret;
}
static int put_ushort(unsigned long arg, unsigned short val)
@@ -553,12 +562,10 @@ int blkdev_ioctl(struct block_device *bd
if (copy_from_user(range, (void __user *)arg, 24))
return -EFAULT;
- ret = blk_ioctl_copy(bdev, range[0], range[1], range[2]);
- if (!ret) {
- range[3] = range[2];
- if (copy_to_user((void __user *)(arg + 24), &range[3], 8))
- return -EFAULT;
- }
+ ret = blk_ioctl_copy(bdev, range[0], range[1], range[2], &range[3]);
+
+ if (copy_to_user((void __user *)(arg + 24), &range[3], 8))
+ return -EFAULT;
return ret;
}
The block layer creates two bios for each copy operation. The bios travel
independently through the storage stack and they are paired at the block
device.
There is a theoretical problem with this - the block device stack only
guarantees forward progress for a single bio. When two bios are sent, it
is possible (though very unlikely) that the first bio exhausts some
mempool and the second bio waits until there is free space in the mempool
(and thus it waits until the first bio finishes).
To avoid this deadlock, we introduce a timer. If the two bios are not
paired at the physical block device within 10 seconds, the copy operation
is aborted and the bio that waits to be paired is released with an error.
Note that there is no guarantee that any XCOPY operation succeed, so
aborting an operation with an error shouldn't cause any problems - the
caller is supposed to perform the copy manually if XCOPY fails.
Signed-off-by: Mikulas Patocka <[email protected]>
---
block/blk-lib.c | 31 +++++++++++++++++++++++++++++++
include/linux/blk_types.h | 2 ++
2 files changed, 33 insertions(+)
Index: linux-4.3/block/blk-lib.c
===================================================================
--- linux-4.3.orig/block/blk-lib.c 2015-11-02 18:43:06.000000000 +0100
+++ linux-4.3/block/blk-lib.c 2015-11-02 18:43:10.000000000 +0100
@@ -300,6 +300,34 @@ int blkdev_issue_zeroout(struct block_de
}
EXPORT_SYMBOL(blkdev_issue_zeroout);
+#define BLK_COPY_TIMEOUT (10 * HZ)
+
+static void blk_copy_timeout(unsigned long bc_)
+{
+ struct bio_copy *bc = (struct bio_copy *)bc_;
+ struct bio *bio0 = NULL, *bio1 = NULL;
+
+ WARN_ON(!irqs_disabled());
+
+ spin_lock(&bc->spinlock); /* the timer is IRQSAFE */
+ if (bc->error == 1) {
+ bc->error = -ETIMEDOUT;
+ bio0 = bc->pair[0];
+ bio1 = bc->pair[1];
+ bc->pair[0] = bc->pair[1] = NULL;
+ }
+ spin_unlock(&bc->spinlock);
+
+ if (bio0) {
+ bio0->bi_error = -ETIMEDOUT;
+ bio_endio(bio0);
+ }
+ if (bio1) {
+ bio1->bi_error = -ETIMEDOUT;
+ bio_endio(bio1);
+ }
+}
+
static void bio_copy_end_io(struct bio *bio)
{
struct bio_copy *bc = bio->bi_copy;
@@ -335,6 +363,7 @@ static void bio_copy_end_io(struct bio *
} while (unlikely(atomic64_cmpxchg(bc->first_error,
first_error, bc->offset) != first_error));
}
+ del_timer_sync(&bc->timer);
kfree(bc);
if (atomic_dec_and_test(&bb->done))
complete(bb->wait);
@@ -425,6 +454,8 @@ int blkdev_issue_copy(struct block_devic
bc->first_error = &first_error;
bc->offset = offset;
spin_lock_init(&bc->spinlock);
+ __setup_timer(&bc->timer, blk_copy_timeout, (unsigned long)bc, TIMER_IRQSAFE);
+ mod_timer(&bc->timer, jiffies + BLK_COPY_TIMEOUT);
read_bio->bi_iter.bi_sector = src_sector;
read_bio->bi_iter.bi_size = chunk << 9;
Index: linux-4.3/include/linux/blk_types.h
===================================================================
--- linux-4.3.orig/include/linux/blk_types.h 2015-11-02 18:43:06.000000000 +0100
+++ linux-4.3/include/linux/blk_types.h 2015-11-02 18:43:10.000000000 +0100
@@ -6,6 +6,7 @@
#define __LINUX_BLK_TYPES_H
#include <linux/types.h>
+#include <linux/timer.h>
struct bio_set;
struct bio;
@@ -52,6 +53,7 @@ struct bio_copy {
atomic64_t *first_error;
sector_t offset;
spinlock_t spinlock;
+ struct timer_list timer;
};
/*
In dm-snapshot target there may be large number of copy requests in
progress. If every pending copy request consumed a process context, it
would put too much load on the system.
To avoid this load, we need asynchronous notification when copy finishes -
we can pass a callback to the function blkdev_issue_copy, if the callback
is non-NULL, blkdev_issue_copy exits when it submits all the copy bios and
the callback is called when the copy operation finishes.
With the callback mechanism, there can be large number of in-progress copy
requests and we do not need process context for each of them.
Signed-off-by: Mikulas Patocka <[email protected]>
---
block/blk-lib.c | 148 +++++++++++++++++++++++++++++++++-------------
block/ioctl.c | 2
include/linux/blk_types.h | 5 -
include/linux/blkdev.h | 2
4 files changed, 112 insertions(+), 45 deletions(-)
Index: linux-4.4-rc4/block/blk-lib.c
===================================================================
--- linux-4.4-rc4.orig/block/blk-lib.c 2015-12-10 17:04:45.000000000 +0100
+++ linux-4.4-rc4/block/blk-lib.c 2015-12-10 17:04:46.000000000 +0100
@@ -300,6 +300,17 @@ int blkdev_issue_zeroout(struct block_de
}
EXPORT_SYMBOL(blkdev_issue_zeroout);
+struct bio_copy_batch {
+ atomic_long_t done;
+ int async_error;
+ int sync_error;
+ sector_t sync_copied;
+ atomic64_t first_error;
+ void (*callback)(void *data, int error);
+ void *data;
+ sector_t *copied;
+};
+
#define BLK_COPY_TIMEOUT (10 * HZ)
static void blk_copy_timeout(unsigned long bc_)
@@ -328,6 +339,18 @@ static void blk_copy_timeout(unsigned lo
}
}
+static void blk_copy_batch_finish(struct bio_copy_batch *batch)
+{
+ void (*fn)(void *, int) = batch->callback;
+ void *data = batch->data;
+ int error = unlikely(batch->sync_error) ? batch->sync_error : batch->async_error;
+ if (batch->copied)
+ *batch->copied = min(batch->sync_copied, (sector_t)atomic64_read(&batch->first_error));
+ kfree(batch);
+ if (fn)
+ fn(data, error);
+}
+
static void bio_copy_end_io(struct bio *bio)
{
struct bio_copy *bc = bio->bi_copy;
@@ -351,25 +374,37 @@ static void bio_copy_end_io(struct bio *
}
bio_put(bio);
if (atomic_dec_and_test(&bc->in_flight)) {
- struct bio_batch *bb = bc->private;
+ struct bio_copy_batch *batch = bc->batch;
if (unlikely(bc->error < 0)) {
u64 first_error;
- if (!ACCESS_ONCE(bb->error))
- ACCESS_ONCE(bb->error) = bc->error;
+ if (!ACCESS_ONCE(batch->async_error))
+ ACCESS_ONCE(batch->async_error) = bc->error;
do {
- first_error = atomic64_read(bc->first_error);
+ first_error = atomic64_read(&batch->first_error);
if (bc->offset >= first_error)
break;
- } while (unlikely(atomic64_cmpxchg(bc->first_error,
+ } while (unlikely(atomic64_cmpxchg(&batch->first_error,
first_error, bc->offset) != first_error));
}
del_timer_sync(&bc->timer);
kfree(bc);
- if (atomic_dec_and_test(&bb->done))
- complete(bb->wait);
+ if (atomic_long_dec_and_test(&batch->done))
+ blk_copy_batch_finish(batch);
}
}
+struct bio_copy_completion {
+ struct completion wait;
+ int error;
+};
+
+static void bio_copy_sync_callback(void *ptr, int error)
+{
+ struct bio_copy_completion *comp = ptr;
+ comp->error = error;
+ complete(&comp->wait);
+}
+
/**
* blkdev_issue_copy - queue a copy same operation
* @src_bdev: source blockdev
@@ -384,57 +419,83 @@ static void bio_copy_end_io(struct bio *
*/
int blkdev_issue_copy(struct block_device *src_bdev, sector_t src_sector,
struct block_device *dst_bdev, sector_t dst_sector,
- sector_t nr_sects, gfp_t gfp_mask, sector_t *copied)
+ sector_t nr_sects, gfp_t gfp_mask,
+ void (*callback)(void *, int), void *data,
+ sector_t *copied)
{
DECLARE_COMPLETION_ONSTACK(wait);
struct request_queue *sq = bdev_get_queue(src_bdev);
struct request_queue *dq = bdev_get_queue(dst_bdev);
unsigned int max_copy_sectors;
- struct bio_batch bb;
- int ret = 0;
- atomic64_t first_error = ATOMIC64_INIT(nr_sects);
- sector_t offset = 0;
+ int ret;
+ struct bio_copy_batch *batch;
+ struct bio_copy_completion comp;
if (copied)
*copied = 0;
- if (!sq || !dq)
- return -ENXIO;
+ if (!sq || !dq) {
+ ret = -ENXIO;
+ goto end_callback;
+ }
max_copy_sectors = min(sq->limits.max_copy_sectors,
dq->limits.max_copy_sectors);
- if (max_copy_sectors == 0)
- return -EOPNOTSUPP;
+ if (max_copy_sectors == 0) {
+ ret = -EOPNOTSUPP;
+ goto end_callback;
+ }
if (src_sector + nr_sects < src_sector ||
- dst_sector + nr_sects < dst_sector)
- return -EINVAL;
+ dst_sector + nr_sects < dst_sector) {
+ ret = -EINVAL;
+ goto end_callback;
+ }
/* Do not support overlapping copies */
if (src_bdev == dst_bdev &&
- abs((u64)dst_sector - (u64)src_sector) < nr_sects)
- return -EOPNOTSUPP;
+ abs((u64)dst_sector - (u64)src_sector) < nr_sects) {
+ ret = -EOPNOTSUPP;
+ goto end_callback;
+ }
- atomic_set(&bb.done, 1);
- bb.error = 0;
- bb.wait = &wait;
+ batch = kmalloc(sizeof(struct bio_copy_batch), gfp_mask);
+ if (!batch) {
+ ret = -ENOMEM;
+ goto end_callback;
+ }
+
+ batch->done = (atomic_long_t)ATOMIC_LONG_INIT(1);
+ batch->async_error = 0;
+ batch->sync_error = 0;
+ batch->sync_copied = 0;
+ batch->first_error = (atomic64_t)ATOMIC64_INIT(nr_sects);
+ batch->copied = copied;
+ if (callback) {
+ batch->callback = callback;
+ batch->data = data;
+ } else {
+ comp.wait = COMPLETION_INITIALIZER_ONSTACK(comp.wait);
+ batch->callback = bio_copy_sync_callback;
+ batch->data = ∁
+ }
- while (nr_sects && !ACCESS_ONCE(bb.error)) {
+ while (nr_sects && !ACCESS_ONCE(batch->async_error)) {
struct bio *read_bio, *write_bio;
struct bio_copy *bc;
unsigned chunk = (unsigned)min(nr_sects, (sector_t)max_copy_sectors);
bc = kmalloc(sizeof(struct bio_copy), gfp_mask);
if (!bc) {
- ret = -ENOMEM;
+ batch->sync_error = -ENOMEM;
break;
}
read_bio = bio_alloc(gfp_mask, 1);
if (!read_bio) {
kfree(bc);
- ret = -ENOMEM;
+ batch->sync_error = -ENOMEM;
break;
}
@@ -442,7 +503,7 @@ int blkdev_issue_copy(struct block_devic
if (!write_bio) {
bio_put(read_bio);
kfree(bc);
- ret = -ENOMEM;
+ batch->sync_error = -ENOMEM;
break;
}
@@ -450,9 +511,8 @@ int blkdev_issue_copy(struct block_devic
bc->error = 1;
bc->pair[0] = NULL;
bc->pair[1] = NULL;
- bc->private = &bb;
- bc->first_error = &first_error;
- bc->offset = offset;
+ bc->batch = batch;
+ bc->offset = batch->sync_copied;
spin_lock_init(&bc->spinlock);
__setup_timer(&bc->timer, blk_copy_timeout, (unsigned long)bc, TIMER_IRQSAFE);
mod_timer(&bc->timer, jiffies + BLK_COPY_TIMEOUT);
@@ -469,27 +529,33 @@ int blkdev_issue_copy(struct block_devic
write_bio->bi_bdev = dst_bdev;
write_bio->bi_copy = bc;
- atomic_inc(&bb.done);
+ atomic_long_inc(&batch->done);
submit_bio(READ | REQ_COPY, read_bio);
submit_bio(WRITE | REQ_COPY, write_bio);
src_sector += chunk;
dst_sector += chunk;
nr_sects -= chunk;
- offset += chunk;
+ batch->sync_copied += chunk;
}
- /* Wait for bios in-flight */
- if (!atomic_dec_and_test(&bb.done))
- wait_for_completion_io(&wait);
+ if (atomic_long_dec_and_test(&batch->done))
+ blk_copy_batch_finish(batch);
- if (copied)
- *copied = min((sector_t)atomic64_read(&first_error), offset);
-
- if (likely(!ret))
- ret = bb.error;
+ if (callback) {
+ return 0;
+ } else {
+ wait_for_completion_io(&comp.wait);
+ return comp.error;
+ }
- return ret;
+end_callback:
+ if (callback) {
+ callback(data, ret);
+ return 0;
+ } else {
+ return ret;
+ }
}
EXPORT_SYMBOL(blkdev_issue_copy);
Index: linux-4.4-rc4/include/linux/blk_types.h
===================================================================
--- linux-4.4-rc4.orig/include/linux/blk_types.h 2015-12-10 17:04:45.000000000 +0100
+++ linux-4.4-rc4/include/linux/blk_types.h 2015-12-10 17:04:46.000000000 +0100
@@ -40,6 +40,8 @@ struct bvec_iter {
current bvec */
};
+struct bio_copy_batch;
+
struct bio_copy {
/*
* error == 1 - bios are waiting to be paired
@@ -49,8 +51,7 @@ struct bio_copy {
int error;
atomic_t in_flight;
struct bio *pair[2];
- void *private;
- atomic64_t *first_error;
+ struct bio_copy_batch *batch;
sector_t offset;
spinlock_t spinlock;
struct timer_list timer;
Index: linux-4.4-rc4/include/linux/blkdev.h
===================================================================
--- linux-4.4-rc4.orig/include/linux/blkdev.h 2015-12-10 17:04:40.000000000 +0100
+++ linux-4.4-rc4/include/linux/blkdev.h 2015-12-10 17:04:46.000000000 +0100
@@ -1142,7 +1142,7 @@ extern int blkdev_issue_write_same(struc
sector_t nr_sects, gfp_t gfp_mask, struct page *page);
extern int blkdev_issue_copy(struct block_device *, sector_t,
struct block_device *, sector_t, sector_t, gfp_t,
- sector_t *);
+ void (*)(void *, int), void *, sector_t *);
extern int blkdev_issue_zeroout(struct block_device *bdev, sector_t sector,
sector_t nr_sects, gfp_t gfp_mask, bool discard);
static inline int sb_issue_discard(struct super_block *sb, sector_t block,
Index: linux-4.4-rc4/block/ioctl.c
===================================================================
--- linux-4.4-rc4.orig/block/ioctl.c 2015-12-10 17:04:40.000000000 +0100
+++ linux-4.4-rc4/block/ioctl.c 2015-12-10 17:04:46.000000000 +0100
@@ -276,7 +276,7 @@ static int blk_ioctl_copy(struct block_d
return -EINVAL;
ret = blkdev_issue_copy(bdev, src_offset, bdev, dst_offset, len,
- GFP_KERNEL, &copied_sec);
+ GFP_KERNEL, NULL, NULL, &copied_sec);
*copied = (uint64_t)copied_sec << 9;
This patch suppresses error messages when copying between two arrays that
support XCOPY each, but that cannot copy data between each other.
Signed-off-by: Mikulas Patocka <[email protected]>
---
drivers/scsi/sd.c | 14 ++++++++++++++
1 file changed, 14 insertions(+)
Index: linux-4.4-rc4/drivers/scsi/sd.c
===================================================================
--- linux-4.4-rc4.orig/drivers/scsi/sd.c 2015-12-07 16:58:48.000000000 +0100
+++ linux-4.4-rc4/drivers/scsi/sd.c 2015-12-07 16:59:09.000000000 +0100
@@ -1995,6 +1995,20 @@ static int sd_done(struct scsi_cmnd *SCp
req->cmd_flags |= REQ_QUIET;
}
}
+ } else if (sshdr.asc == 0x26) {
+ switch (op) {
+ /*
+ * Copying between two arrays that support XCOPY, but
+ * cannot access each other.
+ */
+ case EXTENDED_COPY:
+ if ((SCpnt->cmnd[1] & 0x1f) == 0) {
+ good_bytes = 0;
+ req->__data_len = blk_rq_bytes(req);
+ req->cmd_flags |= REQ_QUIET;
+ }
+ break;
+ }
}
break;
default:
If xcopy between two devices fails, it is pointless to send more xcopy
command between there two devices because they take time and they will
likely also fail.
This patch keeps a cache of (source_device,destination_device) pairs where
copying failed and makes sure that no xcopy command is sooner than 30
seconds after the last failure.
Signed-off-by: Mikulas Patocka <[email protected]>
---
drivers/scsi/sd.c | 37 +++++++++++++++++++++++++++++++++++++
1 file changed, 37 insertions(+)
Index: linux-4.4-rc4/drivers/scsi/sd.c
===================================================================
--- linux-4.4-rc4.orig/drivers/scsi/sd.c 2015-12-07 16:59:09.000000000 +0100
+++ linux-4.4-rc4/drivers/scsi/sd.c 2015-12-07 16:59:12.000000000 +0100
@@ -939,6 +939,26 @@ static void sd_config_copy(struct scsi_d
(logical_block_size >> 9));
}
+#define SD_COPY_DISABLED_CACHE_TIME (HZ * 30)
+#define SD_COPY_DISABLED_CACHE_HASH_BITS 6
+#define SD_COPY_DISABLED_CACHE_HASH (1 << SD_COPY_DISABLED_CACHE_HASH_BITS)
+
+struct sd_copy_disabled_cache_entry {
+ struct scsi_device *src;
+ struct scsi_device *dst;
+ unsigned long jiffies;
+};
+
+static struct sd_copy_disabled_cache_entry sd_copy_disabled_cache[SD_COPY_DISABLED_CACHE_HASH];
+
+static struct sd_copy_disabled_cache_entry *sd_copy_disabled_cache_hash(
+ struct scsi_device *src, struct scsi_device *dst)
+{
+ return &sd_copy_disabled_cache[
+ hash_long((unsigned long)src + ((unsigned long)dst >> 1), SD_COPY_DISABLED_CACHE_HASH_BITS)
+ ];
+}
+
static int sd_setup_copy_cmnd(struct scsi_cmnd *cmd)
{
struct request *rq = cmd->request;
@@ -951,6 +971,7 @@ static int sd_setup_copy_cmnd(struct scs
struct bio *bio = rq->bio;
struct page *page;
unsigned char *buf;
+ struct sd_copy_disabled_cache_entry *e;
dst_sdp = scsi_disk(rq->rq_disk)->device;
dst_queue = rq->rq_disk->queue;
@@ -970,6 +991,12 @@ static int sd_setup_copy_cmnd(struct scs
if (src_sdp->sector_size != dst_sdp->sector_size)
return BLKPREP_KILL;
+ /* The copy failed in the past, so do not retry it for some time */
+ e = sd_copy_disabled_cache_hash(src_sdp, dst_sdp);
+ if (unlikely(jiffies - ACCESS_ONCE(e->jiffies) < SD_COPY_DISABLED_CACHE_TIME) &&
+ likely(ACCESS_ONCE(e->src) == src_sdp) && likely(ACCESS_ONCE(e->dst) == dst_sdp))
+ return BLKPREP_KILL;
+
dst_lba = blk_rq_pos(rq) >> (ilog2(dst_sdp->sector_size) - 9);
src_lba = bio->bi_copy->pair[0]->bi_iter.bi_sector >> (ilog2(src_sdp->sector_size) - 9);
nr_blocks = blk_rq_sectors(rq) >> (ilog2(dst_sdp->sector_size) - 9);
@@ -2003,6 +2030,16 @@ static int sd_done(struct scsi_cmnd *SCp
*/
case EXTENDED_COPY:
if ((SCpnt->cmnd[1] & 0x1f) == 0) {
+ struct sd_copy_disabled_cache_entry *e;
+ struct scsi_device *src_sdp, *dst_sdp;
+
+ dst_sdp = sdkp->device;
+ src_sdp = scsi_disk(req->bio->bi_copy->pair[0]->bi_bdev->bd_disk)->device;
+ e = sd_copy_disabled_cache_hash(src_sdp, dst_sdp);
+ ACCESS_ONCE(e->src) = src_sdp;
+ ACCESS_ONCE(e->dst) = dst_sdp;
+ ACCESS_ONCE(e->jiffies) = jiffies;
+
good_bytes = 0;
req->__data_len = blk_rq_bytes(req);
req->cmd_flags |= REQ_QUIET;
There is no way to split copy requests, so the creator of the requests
(the function blkdev_issue_copy) must make requests with proper size.
Device mapper splits the requests at a boundary between targets or at
a boundary specified by each target driver. We must make sure that the
copy requets do not cross these boundaries.
This patch introduces a new queue limit "copy_boundary", it is log2 of the
boundary in sectors that the request must not cross. Device mapper will
use this limit to propagate its requirements through the device stack.
Signed-off-by: Mikulas Patocka <[email protected]>
---
block/blk-lib.c | 21 ++++++++++++++++++++-
block/blk-settings.c | 17 +++++++++++++++++
block/blk-sysfs.c | 13 +++++++++++++
include/linux/blkdev.h | 1 +
4 files changed, 51 insertions(+), 1 deletion(-)
Index: linux-4.4-rc4/block/blk-settings.c
===================================================================
--- linux-4.4-rc4.orig/block/blk-settings.c 2015-12-10 17:04:30.000000000 +0100
+++ linux-4.4-rc4/block/blk-settings.c 2015-12-10 17:04:54.000000000 +0100
@@ -96,6 +96,7 @@ void blk_set_default_limits(struct queue
lim->chunk_sectors = 0;
lim->max_write_same_sectors = 0;
lim->max_copy_sectors = 0;
+ lim->copy_boundary = 63;
lim->max_discard_sectors = 0;
lim->max_hw_discard_sectors = 0;
lim->discard_granularity = 0;
@@ -311,6 +312,18 @@ void blk_queue_max_copy_sectors(struct r
EXPORT_SYMBOL(blk_queue_max_copy_sectors);
/**
+ * blk_queue_copy_boundary - set a boundary for copy operations. No copy
+ * operation may cross the boundary
+ * @q: the request queue for the device
+ * @copy_boundary: log2 of the copy boundary in sectors
+ **/
+void blk_queue_copy_boundary(struct request_queue *q,
+ unsigned char copy_boundary)
+{
+ q->limits.copy_boundary = copy_boundary;
+}
+
+/**
* blk_queue_max_segments - set max hw segments for a request for this queue
* @q: the request queue for the device
* @max_segments: max number of segments
@@ -552,6 +565,10 @@ int blk_stack_limits(struct queue_limits
t->max_segment_size = min_not_zero(t->max_segment_size,
b->max_segment_size);
+ t->copy_boundary = min(t->copy_boundary, b->copy_boundary);
+ if (start)
+ t->copy_boundary = min(t->copy_boundary, (unsigned char)__ffs64(start));
+
t->misaligned |= b->misaligned;
alignment = queue_limit_alignment_offset(b, start);
Index: linux-4.4-rc4/block/blk-sysfs.c
===================================================================
--- linux-4.4-rc4.orig/block/blk-sysfs.c 2015-12-10 17:04:30.000000000 +0100
+++ linux-4.4-rc4/block/blk-sysfs.c 2015-12-10 17:04:54.000000000 +0100
@@ -199,6 +199,13 @@ static ssize_t queue_copy_max_show(struc
(unsigned long long)q->limits.max_copy_sectors << 9);
}
+static ssize_t queue_copy_boundary_show(struct request_queue *q, char *page)
+{
+ return sprintf(page, "%llu\n",
+ !q->limits.max_copy_sectors || q->limits.copy_boundary == 63 ?
+ 0ULL : 512ULL << q->limits.copy_boundary);
+}
+
static ssize_t
queue_max_sectors_store(struct request_queue *q, const char *page, size_t count)
{
@@ -453,6 +460,11 @@ static struct queue_sysfs_entry queue_co
.show = queue_copy_max_show,
};
+static struct queue_sysfs_entry queue_copy_boundary_entry = {
+ .attr = {.name = "copy_boundary_bytes", .mode = S_IRUGO },
+ .show = queue_copy_boundary_show,
+};
+
static struct queue_sysfs_entry queue_nonrot_entry = {
.attr = {.name = "rotational", .mode = S_IRUGO | S_IWUSR },
.show = queue_show_nonrot,
@@ -509,6 +521,7 @@ static struct attribute *default_attrs[]
&queue_discard_zeroes_data_entry.attr,
&queue_write_same_max_entry.attr,
&queue_copy_max_entry.attr,
+ &queue_copy_boundary_entry.attr,
&queue_nonrot_entry.attr,
&queue_nomerges_entry.attr,
&queue_rq_affinity_entry.attr,
Index: linux-4.4-rc4/include/linux/blkdev.h
===================================================================
--- linux-4.4-rc4.orig/include/linux/blkdev.h 2015-12-10 17:04:46.000000000 +0100
+++ linux-4.4-rc4/include/linux/blkdev.h 2015-12-10 17:04:54.000000000 +0100
@@ -273,6 +273,7 @@ struct queue_limits {
unsigned short max_segments;
unsigned short max_integrity_segments;
+ unsigned char copy_boundary;
unsigned char misaligned;
unsigned char discard_misaligned;
unsigned char cluster;
Index: linux-4.4-rc4/block/blk-lib.c
===================================================================
--- linux-4.4-rc4.orig/block/blk-lib.c 2015-12-10 17:04:46.000000000 +0100
+++ linux-4.4-rc4/block/blk-lib.c 2015-12-10 17:04:54.000000000 +0100
@@ -447,6 +447,12 @@ int blkdev_issue_copy(struct block_devic
goto end_callback;
}
+ /*
+ * TODO: if copy_boundary is too small, we'd better reject this
+ * request and let the caller perform copy manually. The threshold
+ * should be determined experimentally.
+ */
+
if (src_sector + nr_sects < src_sector ||
dst_sector + nr_sects < dst_sector) {
ret = -EINVAL;
@@ -484,7 +490,20 @@ int blkdev_issue_copy(struct block_devic
while (nr_sects && !ACCESS_ONCE(batch->async_error)) {
struct bio *read_bio, *write_bio;
struct bio_copy *bc;
- unsigned chunk = (unsigned)min(nr_sects, (sector_t)max_copy_sectors);
+ unsigned chunk;
+ u64 boundary, max_io;
+
+ chunk = (unsigned)min(nr_sects, (sector_t)max_copy_sectors);
+
+ boundary = 1ULL << sq->limits.copy_boundary;
+ max_io = boundary - (src_sector & (boundary - 1));
+ if (chunk > max_io)
+ chunk = max_io;
+
+ boundary = 1ULL << dq->limits.copy_boundary;
+ max_io = boundary - (dst_sector & (boundary - 1));
+ if (chunk > max_io)
+ chunk = max_io;
bc = kmalloc(sizeof(struct bio_copy), gfp_mask);
if (!bc) {
This patch implements basic copy support for device mapper core.
Individual targets can enable copy support by setting ti->copy_supported.
Device mapper device advertises copy support if at least one target
supports copy and for this target, at least one underlying device supports
copy.
Signed-off-by: Mikulas Patocka <[email protected]>
---
drivers/md/dm-table.c | 16 ++++++++++++++++
drivers/md/dm.c | 27 +++++++++++++++++++++++++++
include/linux/device-mapper.h | 5 +++++
3 files changed, 48 insertions(+)
Index: linux-4.4-rc4/drivers/md/dm.c
===================================================================
--- linux-4.4-rc4.orig/drivers/md/dm.c 2015-12-10 17:04:05.000000000 +0100
+++ linux-4.4-rc4/drivers/md/dm.c 2015-12-10 17:05:48.000000000 +0100
@@ -1679,6 +1679,31 @@ static int __send_write_same(struct clon
return __send_changing_extent_only(ci, get_num_write_same_bios, NULL);
}
+static int __send_copy(struct clone_info *ci)
+{
+ struct dm_target *ti;
+ sector_t bound;
+
+ ti = dm_table_find_target(ci->map, ci->sector);
+ if (!dm_target_is_valid(ti))
+ return -EIO;
+
+ if (!ti->copy_supported)
+ return -EOPNOTSUPP;
+
+ bound = max_io_len(ci->sector, ti);
+
+ if (unlikely(ci->sector_count > bound))
+ return -EOPNOTSUPP;
+
+ __clone_and_map_simple_bio(ci, ti, 0, NULL);
+
+ ci->sector += ci->sector_count;
+ ci->sector_count = 0;
+
+ return 0;
+}
+
/*
* Select the correct strategy for processing a non-flush bio.
*/
@@ -1692,6 +1717,8 @@ static int __split_and_process_non_flush
return __send_discard(ci);
else if (unlikely(bio->bi_rw & REQ_WRITE_SAME))
return __send_write_same(ci);
+ else if (unlikely(bio->bi_rw & REQ_COPY))
+ return __send_copy(ci);
ti = dm_table_find_target(ci->map, ci->sector);
if (!dm_target_is_valid(ti))
Index: linux-4.4-rc4/include/linux/device-mapper.h
===================================================================
--- linux-4.4-rc4.orig/include/linux/device-mapper.h 2015-12-10 17:04:05.000000000 +0100
+++ linux-4.4-rc4/include/linux/device-mapper.h 2015-12-10 17:04:59.000000000 +0100
@@ -271,6 +271,11 @@ struct dm_target {
* Set if this target does not return zeroes on discarded blocks.
*/
bool discard_zeroes_data_unsupported:1;
+
+ /*
+ * Set if the target supports XCOPY.
+ */
+ bool copy_supported:1;
};
/* Each target can link one of these into the table */
Index: linux-4.4-rc4/drivers/md/dm-table.c
===================================================================
--- linux-4.4-rc4.orig/drivers/md/dm-table.c 2015-12-10 17:03:51.000000000 +0100
+++ linux-4.4-rc4/drivers/md/dm-table.c 2015-12-10 17:04:59.000000000 +0100
@@ -286,6 +286,11 @@ static int device_area_is_invalid(struct
limits->logical_block_size >> SECTOR_SHIFT;
char b[BDEVNAME_SIZE];
+ if (ti->copy_supported)
+ limits->max_copy_sectors =
+ min_not_zero(limits->max_copy_sectors,
+ bdev_get_queue(bdev)->limits.max_copy_sectors);
+
/*
* Some devices exist without request functions,
* such as loop devices not yet bound to backing files.
@@ -1256,6 +1261,13 @@ int dm_calculate_queue_limits(struct dm_
ti = dm_table_get_target(table, i++);
+ if (ti->begin)
+ ti_limits.copy_boundary = min(ti_limits.copy_boundary,
+ (unsigned char)__ffs64(ti->begin));
+ if (ti->max_io_len)
+ ti_limits.copy_boundary = min(ti_limits.copy_boundary,
+ (unsigned char)__ffs(ti->max_io_len));
+
if (!ti->type->iterate_devices)
goto combine_limits;
@@ -1289,6 +1301,10 @@ combine_limits:
dm_device_name(table->md),
(unsigned long long) ti->begin,
(unsigned long long) ti->len);
+
+ limits->max_copy_sectors =
+ min_not_zero(limits->max_copy_sectors,
+ ti_limits.max_copy_sectors);
}
return validate_hardware_logical_block_alignment(table, limits);
Support copy operation in the linear target.
Signed-off-by: Mikulas Patocka <[email protected]>
---
drivers/md/dm-linear.c | 1 +
1 file changed, 1 insertion(+)
Index: linux-4.3-rc1/drivers/md/dm-linear.c
===================================================================
--- linux-4.3-rc1.orig/drivers/md/dm-linear.c 2015-09-14 16:06:36.000000000 +0200
+++ linux-4.3-rc1/drivers/md/dm-linear.c 2015-09-14 16:22:44.000000000 +0200
@@ -59,6 +59,7 @@ static int linear_ctr(struct dm_target *
ti->num_flush_bios = 1;
ti->num_discard_bios = 1;
ti->num_write_same_bios = 1;
+ ti->copy_supported = 1;
ti->private = lc;
return 0;
Support the copy operation for the stripe target.
In stripe_merge, we verify that the underlying device supports copy. If it
doesn't, we can fail fast without any bio being contructed.
Signed-off-by: Mikulas Patocka <[email protected]>
---
drivers/md/dm-stripe.c | 1 +
1 file changed, 1 insertion(+)
Index: linux-4.3-rc1/drivers/md/dm-stripe.c
===================================================================
--- linux-4.3-rc1.orig/drivers/md/dm-stripe.c 2015-09-14 16:06:36.000000000 +0200
+++ linux-4.3-rc1/drivers/md/dm-stripe.c 2015-09-14 16:23:01.000000000 +0200
@@ -169,6 +169,7 @@ static int stripe_ctr(struct dm_target *
ti->num_flush_bios = stripes;
ti->num_discard_bios = stripes;
ti->num_write_same_bios = stripes;
+ ti->copy_supported = 1;
sc->chunk_size = chunk_size;
if (chunk_size & (chunk_size - 1))
We move some code to a function submit_job. It is needed for the next
patch that calls submit_job from another place.
Signed-off-by: Mikulas Patocka <[email protected]>
---
drivers/md/dm-kcopyd.c | 19 ++++++++++++-------
1 file changed, 12 insertions(+), 7 deletions(-)
Index: linux-3.16-rc5/drivers/md/dm-kcopyd.c
===================================================================
--- linux-3.16-rc5.orig/drivers/md/dm-kcopyd.c 2014-07-14 16:45:23.000000000 +0200
+++ linux-3.16-rc5/drivers/md/dm-kcopyd.c 2014-07-14 17:28:36.000000000 +0200
@@ -698,6 +698,17 @@ static void split_job(struct kcopyd_job
}
}
+static void submit_job(struct kcopyd_job *job)
+{
+ if (job->source.count <= SUB_JOB_SIZE)
+ dispatch_job(job);
+ else {
+ mutex_init(&job->lock);
+ job->progress = 0;
+ split_job(job);
+ }
+}
+
int dm_kcopyd_copy(struct dm_kcopyd_client *kc, struct dm_io_region *from,
unsigned int num_dests, struct dm_io_region *dests,
unsigned int flags, dm_kcopyd_notify_fn fn, void *context)
@@ -746,13 +757,7 @@ int dm_kcopyd_copy(struct dm_kcopyd_clie
job->context = context;
job->master_job = job;
- if (job->source.count <= SUB_JOB_SIZE)
- dispatch_job(job);
- else {
- mutex_init(&job->lock);
- job->progress = 0;
- split_job(job);
- }
+ submit_job(job);
return 0;
}
This patch adds copy offload support to dm-kcopyd. If copy offload fails,
copying is performed using dm-io, just like before.
There is a module parameter "copy_offload" that can be set to enable or
disable this feature. It can be used to test performance of copy offload.
Signed-off-by: Mikulas Patocka <[email protected]>
---
drivers/md/dm-kcopyd.c | 44 ++++++++++++++++++++++++++++++++++++++++----
1 file changed, 40 insertions(+), 4 deletions(-)
Index: linux-3.18-rc3/drivers/md/dm-kcopyd.c
===================================================================
--- linux-3.18-rc3.orig/drivers/md/dm-kcopyd.c 2014-11-05 18:09:23.000000000 +0100
+++ linux-3.18-rc3/drivers/md/dm-kcopyd.c 2014-11-05 18:13:04.000000000 +0100
@@ -96,6 +96,9 @@ static DEFINE_SPINLOCK(throttle_spinlock
*/
#define MAX_SLEEPS 10
+static bool copy_offload = true;
+module_param(copy_offload, bool, S_IRUGO | S_IWUSR);
+
static void io_job_start(struct dm_kcopyd_throttle *t)
{
unsigned throttle, now, difference;
@@ -358,6 +361,8 @@ struct kcopyd_job {
sector_t progress;
struct kcopyd_job *master_job;
+
+ struct work_struct copy_work;
};
static struct kmem_cache *_job_cache;
@@ -709,6 +714,31 @@ static void submit_job(struct kcopyd_job
}
}
+static void copy_offload_work(struct work_struct *work)
+{
+ struct kcopyd_job *job = container_of(work, struct kcopyd_job, copy_work);
+ sector_t copied;
+
+ blkdev_issue_copy(job->source.bdev, job->source.sector,
+ job->dests[0].bdev, job->dests[0].sector,
+ job->source.count,
+ GFP_NOIO | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN,
+ NULL, NULL, &copied);
+
+ job->source.sector += copied;
+ job->source.count -= copied;
+ job->dests[0].sector += copied;
+ job->dests[0].count -= copied;
+
+ submit_job(job);
+}
+
+static void try_copy_offload(struct kcopyd_job *job)
+{
+ INIT_WORK(&job->copy_work, copy_offload_work);
+ queue_work(job->kc->kcopyd_wq, &job->copy_work);
+}
+
int dm_kcopyd_copy(struct dm_kcopyd_client *kc, struct dm_io_region *from,
unsigned int num_dests, struct dm_io_region *dests,
unsigned int flags, dm_kcopyd_notify_fn fn, void *context)
@@ -733,10 +763,20 @@ int dm_kcopyd_copy(struct dm_kcopyd_clie
job->num_dests = num_dests;
memcpy(&job->dests, dests, sizeof(*dests) * num_dests);
+ job->fn = fn;
+ job->context = context;
+ job->master_job = job;
+
if (from) {
job->source = *from;
job->pages = NULL;
job->rw = READ;
+ if (copy_offload && num_dests == 1 &&
+ bdev_copy_offload(job->source.bdev) &&
+ bdev_copy_offload(job->dests[0].bdev)) {
+ try_copy_offload(job);
+ return 0;
+ }
} else {
memset(&job->source, 0, sizeof job->source);
job->source.count = job->dests[0].count;
@@ -753,10 +793,6 @@ int dm_kcopyd_copy(struct dm_kcopyd_clie
}
}
- job->fn = fn;
- job->context = context;
- job->master_job = job;
-
submit_job(job);
return 0;
job->lock is only taken for a finite amount of time and the process
doesn't block while holding it, so change it from mutex to spinlock.
This change is needed for the next patch that makes it possible to call
segment_complete from an interrupt. Taking mutexes inside an interrupt is
not allowed.
Signed-off-by: Mikulas Patocka <[email protected]>
---
drivers/md/dm-kcopyd.c | 10 +++++-----
1 file changed, 5 insertions(+), 5 deletions(-)
Index: linux-3.16-rc5/drivers/md/dm-kcopyd.c
===================================================================
--- linux-3.16-rc5.orig/drivers/md/dm-kcopyd.c 2014-07-15 19:20:34.000000000 +0200
+++ linux-3.16-rc5/drivers/md/dm-kcopyd.c 2014-07-15 19:24:20.000000000 +0200
@@ -21,7 +21,7 @@
#include <linux/slab.h>
#include <linux/vmalloc.h>
#include <linux/workqueue.h>
-#include <linux/mutex.h>
+#include <linux/spinlock.h>
#include <linux/delay.h>
#include <linux/device-mapper.h>
#include <linux/dm-kcopyd.h>
@@ -356,7 +356,7 @@ struct kcopyd_job {
* These fields are only used if the job has been split
* into more manageable parts.
*/
- struct mutex lock;
+ spinlock_t lock;
atomic_t sub_jobs;
sector_t progress;
@@ -629,7 +629,7 @@ static void segment_complete(int read_er
struct kcopyd_job *job = sub_job->master_job;
struct dm_kcopyd_client *kc = job->kc;
- mutex_lock(&job->lock);
+ spin_lock(&job->lock);
/* update the error */
if (read_err)
@@ -653,7 +653,7 @@ static void segment_complete(int read_er
job->progress += count;
}
}
- mutex_unlock(&job->lock);
+ spin_unlock(&job->lock);
if (count) {
int i;
@@ -708,7 +708,7 @@ static void submit_job(struct kcopyd_job
if (job->source.count <= SUB_JOB_SIZE)
dispatch_job(job);
else {
- mutex_init(&job->lock);
+ spin_lock_init(&job->lock);
job->progress = 0;
split_job(job);
}
Change dm kcopyd so that it calls blkdev_issue_copy with an asynchronous
callback. There can be large number of pending kcopyd requests and holding
a process context for each of them may put too much load on the workqueue
subsystem.
This patch changes it so that blkdev_issue_copy returns after it submitted
the requests and copy_offload_callback is called when the copy operation
finishes.
Signed-off-by: Mikulas Patocka <[email protected]>
---
drivers/md/dm-kcopyd.c | 33 ++++++++++++++-------------------
1 file changed, 14 insertions(+), 19 deletions(-)
Index: linux-3.16-rc5/drivers/md/dm-kcopyd.c
===================================================================
--- linux-3.16-rc5.orig/drivers/md/dm-kcopyd.c 2014-07-15 19:24:20.000000000 +0200
+++ linux-3.16-rc5/drivers/md/dm-kcopyd.c 2014-07-15 19:24:54.000000000 +0200
@@ -361,8 +361,6 @@ struct kcopyd_job {
sector_t progress;
struct kcopyd_job *master_job;
-
- struct work_struct copy_work;
};
static struct kmem_cache *_job_cache;
@@ -628,8 +626,9 @@ static void segment_complete(int read_er
struct kcopyd_job *sub_job = (struct kcopyd_job *) context;
struct kcopyd_job *job = sub_job->master_job;
struct dm_kcopyd_client *kc = job->kc;
+ unsigned long flags;
- spin_lock(&job->lock);
+ spin_lock_irqsave(&job->lock, flags);
/* update the error */
if (read_err)
@@ -653,7 +652,7 @@ static void segment_complete(int read_er
job->progress += count;
}
}
- spin_unlock(&job->lock);
+ spin_unlock_irqrestore(&job->lock, flags);
if (count) {
int i;
@@ -714,29 +713,25 @@ static void submit_job(struct kcopyd_job
}
}
-static void copy_offload_work(struct work_struct *work)
+static void copy_offload_callback(void *ptr, int error)
{
- struct kcopyd_job *job = container_of(work, struct kcopyd_job, copy_work);
- sector_t copied;
+ struct kcopyd_job *job = ptr;
- blkdev_issue_copy(job->source.bdev, job->source.sector,
- job->dests[0].bdev, job->dests[0].sector,
- job->source.count,
- GFP_NOIO | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN,
- NULL, NULL, &copied);
-
- job->source.sector += copied;
- job->source.count -= copied;
- job->dests[0].sector += copied;
- job->dests[0].count -= copied;
+ job->source.sector += job->progress;
+ job->source.count -= job->progress;
+ job->dests[0].sector += job->progress;
+ job->dests[0].count -= job->progress;
submit_job(job);
}
static void try_copy_offload(struct kcopyd_job *job)
{
- INIT_WORK(&job->copy_work, copy_offload_work);
- queue_work(job->kc->kcopyd_wq, &job->copy_work);
+ blkdev_issue_copy(job->source.bdev, job->source.sector,
+ job->dests[0].bdev, job->dests[0].sector,
+ job->source.count,
+ GFP_NOIO | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN,
+ copy_offload_callback, job, &job->progress);
}
int dm_kcopyd_copy(struct dm_kcopyd_client *kc, struct dm_io_region *from,
>>>>> "Mikulas" == Mikulas Patocka <[email protected]> writes:
Mikulas,
Mikulas> This patch series adds copy offload (the XCOPY command) to the
Mikulas> block layer, SCSI subsystems and device mapper.
Now that the VFS stuff appears to stabilize I agree it's a good time to
revisit all this. I just merged the required VPD patches from Hannes so
those will be in 4.5.
I have a bunch of changes to the SCSI code that I worked on over the
spring/summer based on a feedback from the array vendors after
discussions we started at LSF/MM. Generally speaking, their comments
didn't make things easier, nor prettier :( But your two bio approach is
a requirement to accommodate those needs (token-based copy) so I'll work
on consolidating your changes with mine.
That said, we still need Mike Christie's patches to go in first.
Mike: What's your status? I'm afraid I didn't get a chance to dig very
deep in your series since it coincided with me scrambling to sort out
SCSI for 4.4. Do you think there's a chance we could get your patches in
shape for 4.5? Is there an up-to-date tree I can look at?
--
Martin K. Petersen Oracle Linux Engineering
On 12/10/2015 04:33 PM, Martin K. Petersen wrote:
> Mike: What's your status? I'm afraid I didn't get a chance to dig very
> deep in your series since it coincided with me scrambling to sort out
> SCSI for 4.4. Do you think there's a chance we could get your patches in
> shape for 4.5? Is there an up-to-date tree I can look at?
I just had 2 review comments left:
1. Originally, I had left REQ_FLUSH a flag. Christoph suggested to break
it up into a op and flag:
http://marc.info/?l=linux-scsi&m=144689113106515&w=2
I started this and messed up :) Was going to retry next week.
2. Start REQ_OP_READ off at non-zero to try and shake out code that was
not converted.
There are a several places where we assume reads are zero and writes are
1 for things like indexing in arrays (like blktrace's ddir_act or dm
starts), passing into block functions (like nvme_alloc_request's call of
blk_mq_alloc_request), and if/else's. I am not done fixing all of them
and testing.
Also, the btrfs patch is really large (1000 lines) because that code
base is so large and there were so many places we passed around rw to
through multiple functions. I wanted to try and break it up, so it would
be easier for those guys to review.
I attached a tarball of the patches here.
Patches made against Jens's linux-block tree:
https://git.kernel.org/cgit/linux/kernel/git/axboe/linux-block.git/
for-next branch. Last commit was:
commit 4c93772e14f0b20f38462c395336a6f5c8c7727d
Merge: e1b98c0 8c0b391
Author: Jens Axboe <[email protected]>
Date: Wed Dec 9 10:57:21 2015 -0700
Merge branch 'for-4.5/nvme' into for-next
On 12/10/2015 04:33 PM, Martin K. Petersen wrote:
>>>>>> "Mikulas" == Mikulas Patocka <[email protected]> writes:
>
> Mikulas,
>
> Mikulas> This patch series adds copy offload (the XCOPY command) to the
> Mikulas> block layer, SCSI subsystems and device mapper.
>
> Now that the VFS stuff appears to stabilize I agree it's a good time to
> revisit all this. I just merged the required VPD patches from Hannes so
> those will be in 4.5.
>
> I have a bunch of changes to the SCSI code that I worked on over the
> spring/summer based on a feedback from the array vendors after
> discussions we started at LSF/MM. Generally speaking, their comments
> didn't make things easier, nor prettier :( But your two bio approach is
> a requirement to accommodate those needs (token-based copy) so I'll work
> on consolidating your changes with mine.
>
> That said, we still need Mike Christie's patches to go in first.
>
> Mike: What's your status? I'm afraid I didn't get a chance to dig very
> deep in your series since it coincided with me scrambling to sort out
> SCSI for 4.4. Do you think there's a chance we could get your patches in
> shape for 4.5? Is there an up-to-date tree I can look at?
>
It looks like my original mail did not make it due to an attachment.
I just had 2 review comments left:
1. Originally, I had left REQ_FLUSH a flag. Christoph suggested to break
it up into a op and flag:
http://marc.info/?l=linux-scsi&m=144689113106515&w=2
I started this and messed up Was going to retry next week.
2. Start REQ_OP_READ off at non-zero to try and shake out code that was
not converted.
There are a several places where we assume reads are zero and writes are
1 for things like indexing in arrays (like blktrace's ddir_act or dm
starts), passing into block functions (like nvme_alloc_request's call of
blk_mq_alloc_request), and if/else's. I am not done fixing all of them
and testing.
Ok, 3 comments. One I gave myself:
3. Also, the btrfs patch is really large (1000 lines) because that code
base is so large and there were so many places we passed around rw to
through multiple functions. I wanted to try and break it up, so it would
be easier for those guys to review.
I uploaded a git tree here:
https://github.com/mikechristie/kernel.git
The patches are in the for-next-req-op.
I made this over Jens's for-next branch in his linux-block tree.
On Thu, Dec 10, 2015 at 11:06:04PM -0600, Mike Christie wrote:
>
> 2. Start REQ_OP_READ off at non-zero to try and shake out code that was
> not converted.
>
> There are a several places where we assume reads are zero and writes are
> 1 for things like indexing in arrays (like blktrace's ddir_act or dm
> starts), passing into block functions (like nvme_alloc_request's call of
> blk_mq_alloc_request), and if/else's. I am not done fixing all of them
> and testing.
If this turns out to be too painful we should just skip it. I though it
would be a nie debugging aid, but we don't need to do it at any cost.
On Thu, 10 Dec 2015, Martin K. Petersen wrote:
> >>>>> "Mikulas" == Mikulas Patocka <[email protected]> writes:
>
> Mikulas,
>
> Mikulas> This patch series adds copy offload (the XCOPY command) to the
> Mikulas> block layer, SCSI subsystems and device mapper.
>
> Now that the VFS stuff appears to stabilize I agree it's a good time to
> revisit all this. I just merged the required VPD patches from Hannes so
> those will be in 4.5.
>
> I have a bunch of changes to the SCSI code that I worked on over the
> spring/summer based on a feedback from the array vendors after
> discussions we started at LSF/MM. Generally speaking, their comments
> didn't make things easier, nor prettier :( But your two bio approach is
> a requirement to accommodate those needs (token-based copy) so I'll work
> on consolidating your changes with mine.
Let me know when you make some progress with that.
Is there some software iSCSI implementation that supports token-based
copy? So that I could try to make support for it.
Mikulas
> That said, we still need Mike Christie's patches to go in first.
>
> Mike: What's your status? I'm afraid I didn't get a chance to dig very
> deep in your series since it coincided with me scrambling to sort out
> SCSI for 4.4. Do you think there's a chance we could get your patches in
> shape for 4.5? Is there an up-to-date tree I can look at?
>
> --
> Martin K. Petersen Oracle Linux Engineering
Mikulas
On 15-12-14 08:22 PM, Mikulas Patocka wrote:
>
>
> On Thu, 10 Dec 2015, Martin K. Petersen wrote:
>
>>>>>>> "Mikulas" == Mikulas Patocka <[email protected]> writes:
>>
>> Mikulas,
>>
>> Mikulas> This patch series adds copy offload (the XCOPY command) to the
>> Mikulas> block layer, SCSI subsystems and device mapper.
>>
>> Now that the VFS stuff appears to stabilize I agree it's a good time to
>> revisit all this. I just merged the required VPD patches from Hannes so
>> those will be in 4.5.
>>
>> I have a bunch of changes to the SCSI code that I worked on over the
>> spring/summer based on a feedback from the array vendors after
>> discussions we started at LSF/MM. Generally speaking, their comments
>> didn't make things easier, nor prettier :( But your two bio approach is
>> a requirement to accommodate those needs (token-based copy) so I'll work
>> on consolidating your changes with mine.
>
> Let me know when you make some progress with that.
>
> Is there some software iSCSI implementation that supports token-based
> copy? So that I could try to make support for it.
FreeNAS. See:
http://www.freenas.org/
It has the best ODX implementation (thanks to Alexander Motin) that
I have come across for testing my ddpt based ODX client. It also
supports large parts of the more general xcopy LID1 and LID4.
BTW In SPC-5 xcopy LID1 has been dropped and "xcopy" now means
xcopy LID4 (as found in SPC-4).
Doug Gilbert
>> That said, we still need Mike Christie's patches to go in first.
>>
>> Mike: What's your status? I'm afraid I didn't get a chance to dig very
>> deep in your series since it coincided with me scrambling to sort out
>> SCSI for 4.4. Do you think there's a chance we could get your patches in
>> shape for 4.5? Is there an up-to-date tree I can look at?
>>
>> --
>> Martin K. Petersen Oracle Linux Engineering
>>>>> "Mikulas" == Mikulas Patocka <[email protected]> writes:
Mikulas> Is there some software iSCSI implementation that supports
Mikulas> token-based copy? So that I could try to make support for it.
I did write support for token-based copy but it's in a different branch
from the xcopy stuff. I will have to reconcile those two branches and
then layer your stuff on top.
--
Martin K. Petersen Oracle Linux Engineering