multiple device driver (linux software raid)
DESCRIPTION
Multiple Device Driver (Linux Software RAID). Ted Baker Andy Wang CIS 4930 / COP 5641. The md driver. Provides virtual devices Created from one or more independent underlying devices The basic mechanism to support RAIDs Redundant arrays of inexpensive disks. RAID0 Striping RAID1 - PowerPoint PPT PresentationTRANSCRIPT
Multiple Device Driver (Linux Software RAID)
Ted Baker Andy WangCIS 4930 / COP 5641
The md driver
Provides virtual devices Created from one or more
independent underlying devices The basic mechanism to support
RAIDs Redundant arrays of inexpensive
disks
Common RAID levels RAID0
Striping RAID1
Mirroring RAID4 (> 3 disks)
Striped array with a parity device
RAID5 (> 3 disks) Striped array with
distributed parity RAID6 (> 4 disks)
Striped array with dual redundancy information
Common RAID levels RAID1+0
Striped array of mirrored disks
RAID0+1 Mirroring two
RAID0s RAID5+0
Striped array of RAID5s
RAID5+1 Mirroring two
RAID5s
md pseudo RAID configurations
Linear (catenates multiple disks into a single one)
Multipath A set of different interfaces to the same
device (e.g., multiple disk controllers) Faulty
A layer over a single device into which errors can be injected
RAID Creation> mdadm --create /dev/md0 --level=1 --raid-devices=2
/dev/hd[ac]1
Create /dev/md0 as RAID1 Consisting of /dev/hda1 and
/dev/hdc1
RAID Status
To check the status for RAIDs See /proc/mdstatPersonalities : [raid1]
md0 : active raid1 sda5[0] sdb5[1]
979840 blocks [2/2] [UU]
md1 : active raid1 sda6[2] sdb6[1]
159661888 blocks [2/1] [_U]
[===>.................] recovery = 17.9%
(28697920/159661888) finish=56.4min speed=38656K/sec
unused devices: <none>
md Super Block
Each device in a RAID may have a superblock with various information Level UUID
128 bit identifier that identifies an array
Some RAID Concepts Personality
RAID level Chunk size
Power of two > 4KB
A RAID assigns chunks to disks in a round robin fashion
Stripe A collection of ith
chunk at each disk form a stripe
Parity A chunk
constructed via XORing other chunks
Synchrony
An update may involve both the data block and the parity block
Implications A RAID may be shut down in an
inconsistency state Resynchronization may be required at
startup, in the background Reduced performance
Recovery
If the md driver detects a write error, it immediately disables that device Continues operation on the remaining
devices Starts recreating the content if there
is a spare drive
Recovery
If the md driver detects a read error Overwrites the bad block Read the block again
If fails, treat it as a write error
Recovery is a background process Can be configured via
/proc/sys/dev/raid/speed_limit_min /proc/sys/dev/raid/speed_limit_max
Bitmap Write-Intent Logging
Records which blocks of the array may be out of sync
Speeds up resynchronization Allows a disk to be temporarily
removed and reinserted without causing an enormous recovery cost Can spin down disks for power savings
Bitmap Write-Intent Logging
Can be stored on a separate device
Write-Behind
Certain devices in the array can be flagged as write-mostly
md will not wait for writes to write-behind devices to complete before returning to the file system
Restriping (Reshaping)
Change the number of disks Change the RAID levels Not robust against failures
faulty.cstatic int __init raid_init(void) {
return register_md_personality(&faulty_personality);
}
static void raid_exit(void) {
unregister_md_personality(&faulty_personality);
}
module_init(raid_init);
module_exit(raid_exit);
faulty.cstatic struct mdk_personality faulty_personality = {
.name = "faulty",
.level = LEVEL_FAULTY,
.owner = THIS_MODULE,
.make_request = make_request,
.run = run,
.stop = stop,
.status = status,
.check_reshape = reshape,
.size = faulty_size
};
faulty.cstatic int run(mddev_t *mddev) {
mdk_rdev_t *rdev;
struct list_head *tmp;
int i;
conf_t *conf = kmalloc(sizeof(*conf), GFP_KERNEL);
.../* error handling + zero out conf */
list_for_each_entry(rdev, mddev, same_set)
conf->rdev = rdev;
md_set_array_sectors(mddev, mddev->dev_sectors);
mddev->private = conf;
reshape(mddev);
return 0;
}
typedef struct faulty_conf { int period[Modes]; atomic_t counters[Modes]; sector_t faults[MaxFault]; int modes[MaxFault]; int nfaults; mdk_rdev_t *rdev;} conf_t;
A field in mdk_rdev_t
list head
faulty.cstatic int reshape(mddev_t *mddev) {
int mode = mddev->new_layout & ModeMask;
int count = mddev->new_layout >> ModeShift;
conf_t *conf = mddev->private;
.../* error checks */
if (mode == /* clear something */)
/* clear various counters */
} else if (mode < Modes) {
conf->period[mode] = count;
if (!count) count++;
atomic_set(&conf->counters[mode], count);
} else ...
return 0;
}
Total number of failure modes (e.g., transient write failure mode)
faulty.cstatic int stop(mddev_t *mddev) {
conf_t *conf = (conf_t *)mddev->private;
kfree(conf);
mddev->private = NULL;
return 0;
}
faulty.cstatic int make_request(request_queue_t *q, struct bio *bio) {
mddev_t *mddev = q->queuedata;
conf_t *conf = (conf_t*)mddev->private;
int failit = 0;
if (bio_data_dir(bio) == WRITE) { /* data direction */
.../* misc cases */
/* if a sector failed before, need to stay failed */
if (check_sector(conf, bio->bi_sector, bio->bi_sector +
(bio->bi_size >> 9), WRITE))
failit = 1;
/* if the period (some predefined constant) is reached
for a sector, record the sector and fail it */
if (check_mode(conf, WritePersistent)) {
add_sector(conf, bio->bi_sector, WritePersistent);
failit = 1;
} ...
faulty.c } else { /* failure cases for reads */
...
}
if (failit) {
struct bio *b = bio_clone(bio, GFP_NOIO);
b->bi_bdev = conf->rdev->bdev;
b->bi_private = bio;
b->bi_end_io = faulty_fail;
generic_make_request(b);
return 0;
} else {
bio->bi_bdev = conf->rdev->bdev;
return 1;
}
}
To the queue of this device,
initialized in md.c from the disk device inode
Make bio point to the actual device, and let the main block layer submit the IO and resolve the
recursion
faulty.cstatic int faulty_fail(struct bio *bio, int error) {
struct bio *b = bio->bi_private;
b->bi_size = bio->bi_size;
b->bi_sector = bio->bi_sector;
bio_put(bio);
bio_io_error(b);
}
blk-core.c
A file system eventually calls __generic_make_request()
static inline void __generic_make_request(struct bio *bio) {
...
do {
...
q = bdev_get_queue(bio->bi_bdev);
.../* check errors */
ret = q->make_request_fn(q, bio);
} while (ret);
}
linear.cstatic int __init linear_init(void) {
return register_md_personality(&linear_personality);
}
static void linear_exit (void) {
unregister_md_personality(&linear_personality);
}
module_init(linear_init);
module_exit(linear_exit);
linear.cstatic struct mdk_personality linear_personality = {
.name = "linear",
.level = LEVEL_LINEAR,
.owner = THIS_MODULE,
.make_request = linear_make_request,
.run = linear_run,
.stop = linear_stop,
.status = linear_status, /* for proc */
.hot_add_disk = linear_add,
.size = linear_size,
};
linear.c
typedef struct linear_private_data { sector_t array_sectors; dev_info_t disks[0]; struct rcu_head rcu;} linear_conf_t;
static int linear_run(mddev_t *mddev) {
linear_conf_t *conf;
/* initialize
conf = linear_conf(mddev, mddev->raid_disks);
if (!conf) return 1;
mddev->private = conf;
md_set_array_sectors(mddev, conf->array_sectors;
...
initialize conf->disks[i].end_sector
linear.c ...
/* determines whether two bio can be merged */
/* overrides the default merge_bvec function */
blk_queue_merge_bvec(mddev->queue, linear_mergeable_bvec);
/* queues are first plugged to build up the queue length, then unplugged to release requests to devices */
mddev->queue->unplug_fn = linear_unplug;
/* disable prefetching when the device is congested */
mddev->queue->backing_dev_info.congested_fn
= linear_congested;
mddev->queue->backing_dev_info.congested_data = mddev;
md_integrity_register(mddev);
return 0;
}
linear.cstatic int linear_stop(mddev_t *mddev) {
linear_conf_t *conf = mddev->private;
/* the unplug fn references 'conf‘ */
rcu_barrier();
blk_sync_queue(mddev->queue);
kfree(conf);
return 0;
}
linear.cstatic int linear_make_request(request_queue_t *q,
struct bio *bio) {
const int rw = bio_data_dir(bio);
mddev_t *mddev = q->queuedata;
dev_info_t *tmp_dev;
sector_t start_sector;
.../* check for errors and update statistics */
rcu_read_lock();
tmp_dev = which_dev(mddev, bio->bi_sector);
start_sector = tmp_dev->end_sector – tmp_dev->rdev->sectors;
.../* more error checks */
linear.c if (unlikely(bio->bi_sector + (bio->bi_size >> 9) >
tmp_dev->end_sector)) {
/* This bio crosses a device boundary, so we have to
* split it. */
struct bio_pair *bp;
sector_t end_sector = tmp_dev->end_sector;
rcu_read_unlock();
bp = bio_split(bio, end_sector – bio->bi_sector);
if (linear_make_request(q, &bp->bio1)) /* recursion!?# */
generic_make_request(&bp->bio1);
if (linear_make_request(q, &bp->bio2)) /* recursion#!% */
generic_make_request(&bp->bio2);
bio_pair_release(bp); /* remove bio hazard */
return 0;
}
linear.c bio->bi_bdev = tmp_dev->rdev->bdev;
bio->bi_sector = bio->bi_sector – start_sector +
tmp_dev->rdev->data_offset;
rcu_read_unlock();
return 1;
}
Again, let the main block layer submit the IO and resolve the recursion
Points to the specific device instead of the
linear device
Translates the virtual sector number to the
physical sector number for the specific device