dm-writeboost: Progress Report

From: Akira Hayakawa
Date: Mon Nov 04 2013 - 00:18:20 EST


Hi, DM Guys

Let me share a new progress report.
I am sorry I have been off for weeks.

Writeboost is getting better I believe.

Many progress, please git pull
https://github.com/akiradeveloper/dm-writeboost.git

1. Removing version switch macros
Previously the code included 10 or more version switches
and thus is not appropriate for upstream code of course.
Now I removed them all and Writeboost is not available for
older version of kernel.

2. .postsuspend and .resume implemented
dmsetup suspend and resume works now.

I implemented .postsuspend and .resume for this purpose.
.postsuspend is called after all I/O is submitted to
the device. In .postsuspend I make all the transient data
persistent.

static void writeboost_postsuspend(struct dm_target *ti)
{
int r;

struct wb_device *wb = ti->private;
struct wb_cache *cache = wb->cache;

flush_current_buffer(cache);
IO(blkdev_issue_flush(cache->device->bdev, GFP_NOIO, NULL));
}

.postsuspend is also called before .dtr is called.
I removed the same code from .dtr therefore.

.resume does nothing at all.

3. Blocking up the device thoroughly implemented
How to block up the device has been my annoyance for weeks.
Now, Writeboost can block up the device in I/O error correctly
and elegantly.

What a dead device should satisfy are
1) Returns error on all incoming requests
2) ACK to all the requests accepted
3) To clean up data structures on memory
`dmsetup remove` must be accepted even after marked dead
4) Blocking up must be processed regardless of any timing issues.

The principle behind how Writeboost blocks up a device
is really simple.

1) All I/O to underlying devices are wrapped by IO macro (shown below)
which set flag WB_DEAD (using set_bit) on -EIO returned.
2) If WB_DEAD is set, all I/O to the underlying devices are ignored.
This is equivalent to having /dev/null devices underlying.
Processings on the memory data structures are still
continuing completely agnostic to the fact that
the underlying devices ignore their requests.
This ensures the logic on the data structures never corrupt
and data on the underlying devices never change after marked dead.
3) Some codes are wrapped by LIVE, DEAD or LIVE_DEAD
macro which switches its behavior in accordance with
WB_DEAD bit.

I tested the mechanism by wrapping the underlying devices (like below)
by dm-flakey. For both cases (cache fails or backing fails)
it seems to be working.

sz=`blockdev --getsize ${CACHE}`
dmsetup create cache-flakey --table "0 ${sz} flakey ${CACHE} 0 5 1"
CACHE=/dev/mapper/cache-flakey

sz=`blockdev --getsize ${BACKING}`
dmsetup create backing-flakey --table "0 ${sz} flakey ${BACKING} 0 20 0"
BACKING=/dev/mapper/backing-flakey

3.1 WARN_ONCE introduced
IO macro branches the behavior with the return code from it wrapping procedure.
It only treat -EIO as a trigger to block up the device and others are not.
The problem remains is that I don't know what error
code can return within possibility.
If the error code is not expected, it WARN_ONCE
to ask user to report the error.

/*
* Only -EIO returned is regarded as a signal
* to block up the whole system.
*
* -EOPNOTSUPP could be returned if the target device is
* a virtual device and we request discard.
*
* -ENOMEM could be returned from blkdev_issue_discard (3.12-rc5)
* for example. Waiting for a while can make room for allocation.
*
* For other unknown error codes we ignore them and
* ask the human users to report us.
*/
#define IO(proc) \
do { \
r = 0; \
LIVE(r = proc); \
if (r == -EOPNOTSUPP) { \
r = 0; \
} else if (r == -EIO) { \
set_bit(WB_DEAD, &wb->flags); \
wake_up_all(&wb->dead_wait_queue); \
WBERR("marked as dead"); \
} else if (r == -ENOMEM) { \
schedule_timeout_interruptible(msecs_to_jiffies(1000));\
} else if (r) { \
r = 0;\
WARN_ONCE(1, "PLEASE REPORT!!! I/O FAILED FOR UNKNOWN REASON %d", r); \
} \
} while (r)

3.2 dead state never recover
As Dave's advice, I regard the error as fatal and provide no way to restart.
Removed blockup from message.

3.3 XFS corruption never reproduce
I think the previously discussed error is due to the
misbehavior of Writeboost in blocking up.
For example, some bios are not returned
which may corrupt the XFS AIL lock?

4. On-disk metadata layout changed and
max segment size is reduced to 512KB from 1MB
struct metablock_device was 13 bytes before.
This means that the metadata may straddle two sectors
which causes partial write.
To avoid this, I added padding to make it 16 bytes.
As a result, maximum log size shrinks to 512KB.

5. `flush_supported = true` added
I added `flush_supported = true` to .ctr
dm-cache and dm-thin also do this.
There is no meaning for Writeboost to ignore flush requests.

Below is my wonderings to note and I request for comments
1. Daemon should stop in suspend?
Which level below should a suspended device satisfy?
1) No transient data remains (currently OK)
2) + No any I/O to the underlying devices?
(although it logically irrelevant to the upper layer)
3) + No any process execution is admitted to
reduce power consumption for example?

2. .merge, .hints and .iterate_devices correct?
What is the correct behavior of these methods?

static void writeboost_io_hints(struct dm_target *ti,
struct queue_limits *limits)
{
blk_limits_io_min(limits, 512);
blk_limits_io_opt(limits, 4096);
}

static int writeboost_iterate_devices(struct dm_target *ti,
iterate_devices_callout_fn fn, void *data)
{
struct wb_device *wb = ti->private;
struct dm_dev *orig = wb->device;
sector_t start = 0;
sector_t len = dm_devsize(orig);
return fn(ti, orig, start, len, data);
}

static int writeboost_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
struct bio_vec *biovec, int max_size)
{
struct wb_device *wb = ti->private;
struct dm_dev *device = wb->device;
struct request_queue *q = bdev_get_queue(device->bdev);

if (!q->merge_bvec_fn)
return max_size;

bvm->bi_bdev = device->bdev;
return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
}


Thanks,
Akira
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/