linux-stable.git/drivers/md, branch linux-2.6.17.y

dm: Fix deadlock under high i/o load in raid1 setup.

2006-09-08T21:51:39+00:00

On an nForce4-equipped machine with two SATA disk in raid1 setup using dmraid,
we experienced frequent deadlock of the system under high i/o load.  'cat
/dev/zero > ~/zero' was the most reliable way to reproduce them: Randomly
after a few GB, 'cp' would be left in 'D' state along with kjournald and
kmirrord.  The functions cp and kjournald were blocked in did vary, but
kmirrord's wchan always pointed to 'mempool_alloc()'.  We've seen this pattern
on 2.6.15 and 2.6.17 kernels.  http://lkml.org/lkml/2005/4/20/142 indicates
that this problem has been around even before.

So much for the facts, here's my interpretation: mempool_alloc() first tries
to atomically allocate the requested memory, or falls back to hand out
preallocated chunks from the mempool.  If both fail, it puts the calling
process (kmirrord in this case) on a private waitqueue until somebody refills
the pool.  Where the only 'somebody' is kmirrord itself, so we have a
deadlock.

I worked around this problem by falling back to a (blocking) kmalloc when
before kmirrord would have ended up on the waitqueue.  This defeats part of
the benefits of using the mempool, but at least keeps the system running.  And
it could be done with a two-line change.  Note that mempool_alloc() clears the
GFP_NOIO flag internally, and only uses it to decide whether to wait or return
an error if immediate allocation fails, so the attached patch doesn't change
behaviour in the non-deadlocking case.  Path is against current git
(2.6.18-rc4), but should apply to earlier versions as well.  I've tested on
2.6.15, where this patch makes the difference between random lockup and a
stable system.

Signed-off-by: Daniel Kobras 
Acked-by: Alasdair G Kergon 
Signed-off-by: Andrew Morton 
Signed-off-by: Greg Kroah-Hartman

dm: mirror sector offset fix

2006-09-08T21:51:38+00:00

The device-mapper core does not perform any remapping of bios before passing
them to the targets.  If a particular mapping begins part-way into a device,
targets obtain the sector relative to the start of the mapping by subtracting
ti->begin.

The dm-raid1 target didn't do this everywhere: this patch fixes it, taking
care to subtract ti->begin exactly once for each bio.

[akpm: too late for 2.6.17 - suitable for 2.6.17.x after it has settled]

Signed-off-by: Neil Brown 
Signed-off-by: Alasdair G Kergon 
Signed-off-by: Andrew Morton 
Signed-off-by: Greg Kroah-Hartman

dm: fix block device initialisation

2006-09-08T21:51:38+00:00

In alloc_dev(), we register the device with the block layer and then continue
to initialize the device.  But register_disk() makes the device available to
be opened before we have completed initialising it.

This patch moves the final bits of the initialization above the disk
registration.

[akpm: too late for 2.6.17 - suitable for 2.6.17.x after it has settled]

Signed-off-by: Jeff Mahoney 
Signed-off-by: Alasdair G Kergon 
Signed-off-by: Andrew Morton 
Signed-off-by: Greg Kroah-Hartman

dm: add module ref counting

2006-09-08T21:51:37+00:00

The reference counting on dm-mod is zero if no mapped devices are open.  This
is incorrect, and can lead to an oops if the module is unloaded while mapped
devices exist.

This patch claims a reference to the module whenever a device is created, and
drops it again when the device is freed.

Devices must be removed before dm-mod is unloaded.

[akpm: too late for 2.6.17 - suitable for 2.6.17.x after it has settled]

Signed-off-by: Jeff Mahoney 
Signed-off-by: Alasdair G Kergon 
Signed-off-by: Andrew Morton 
Signed-off-by: Greg Kroah-Hartman

dm: fix mapped device ref counting

2006-09-08T21:51:37+00:00

To avoid races, _minor_lock must be held while changing mapped device
reference counts.

There are a few paths where a mapped_device pointer is returned before a
reference is taken.  This patch fixes them.

[akpm: too late for 2.6.17 - suitable for 2.6.17.x after it has settled]

Signed-off-by: Jeff Mahoney 
Signed-off-by: Alasdair G Kergon 
Signed-off-by: Andrew Morton 
Signed-off-by: Greg Kroah-Hartman

dm: add DMF_FREEING

2006-09-08T21:51:37+00:00

There is a chicken and egg problem between the block layer and dm in which the
gendisk associated with a mapping keeps a reference-less pointer to the
mapped_device.

This patch uses a new flag DMF_FREEING to indicate when the mapped_device is
no longer valid.  This is checked to prevent any attempt to open the device
from succeeding while the device is being destroyed.

[akpm: too late for 2.6.17 - suitable for 2.6.17.x after it has settled]

Signed-off-by: Jeff Mahoney 
Signed-off-by: Alasdair G Kergon 
Signed-off-by: Andrew Morton 
Signed-off-by: Greg Kroah-Hartman

dm: change minor_lock to spinlock

2006-09-08T21:51:37+00:00

While removing a device, another another thread might attempt to resurrect it.

This patch replaces the _minor_lock mutex with a spinlock and uses
atomic_dec_and_lock() to serialize reference counting in dm_put().

[akpm: too late for 2.6.17 - suitable for 2.6.17.x after it has settled]

Signed-off-by: Jeff Mahoney 
Signed-off-by: Alasdair G Kergon 
Signed-off-by: Andrew Morton 
Signed-off-by: Greg Kroah-Hartman

dm: move idr_pre_get

2006-09-08T21:51:37+00:00

idr_pre_get() can sleep while allocating memory.

The next patch will change _minor_lock into a spinlock, so this patch moves
idr_pre_get() outside the lock in preparation.

[akpm: too late for 2.6.17 - suitable for 2.6.17.x after it has settled]

Signed-off-by: Jeff Mahoney 
Signed-off-by: Alasdair G Kergon 
Signed-off-by: Andrew Morton 
Signed-off-by: Greg Kroah-Hartman

dm: fix idr minor allocation

2006-09-08T21:51:37+00:00

One part of the system can attempt to use a mapped device before another has
finished initialising it or while it is being freed.

This patch introduces a place holder value, MINOR_ALLOCED, to mark the minor
as allocated but in a state where it can't be used, such as mid-allocation or
mid-free.  At the end of the initialization, it replaces the place holder with
the pointer to the mapped_device, making it available to the rest of the dm
subsystem.

[akpm: too late for 2.6.17 - suitable for 2.6.17.x after it has settled]

Signed-off-by: Jeff Mahoney 
Signed-off-by: Alasdair G Kergon 
Signed-off-by: Andrew Morton 
Signed-off-by: Greg Kroah-Hartman

dm snapshot: unify chunk_size

2006-09-08T21:51:37+00:00

Persistent snapshots currently store a private copy of the chunk size. 
Userspace also supplies the chunk size when loading a snapshot.  Ensure
consistency by only storing the chunk_size in one place instead of two.


Currently the two sizes will differ if the chunk size supplied by userspace
does not match the chunk size an existing snapshot actually uses.  Amongst
other problems, this causes an incorrect 'percentage full' to be reported.

The patch ensures consistency by only storing the chunk_size in one place,
removing it from struct pstore.  Some initialisation is delayed until the
correct chunk_size is known.  If read_header() discovers that the wrong chunk
size was supplied, the 'area' buffer (which the header already got read into)
is reinitialised to the correct size.

[akpm: too late for 2.6.17 - suitable for 2.6.17.x after it has settled]

Signed-off-by: Alasdair G Kergon 
Signed-off-by: Andrew Morton 
Signed-off-by: Greg Kroah-Hartman