Image files have two types of data: immutable data that describes things like
image size, backing files, etc. and mutable data that includes offset and
reference count tables.
Today, image formats aggressively cache mutable data to improve performance. In
some cases, this happens before a guest even starts. When dealing with live
migration, since a file is open on two machines, the caching of meta data can
lead to data corruption.
This patch addresses this by introducing a mechanism to invalidate any cached
mutable data a block driver may have which is then used by the live migration
code.
NB, this still requires coherent shared storage. Addressing migration without
coherent shared storage (i.e. NFS) requires additional work.
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
cache=unsafe completely ignored bdrv_flush, because flushing the host disk
costs a lot of performance. However, this means that qcow2 images (and
potentially any other format) can lose data even after the guest has issued a
flush if the qemu process crashes/is killed. In case of a host crash, data loss
is certainly expected with cache=unsafe, but if just the qemu process dies this
is a bit too unsafe.
Now that we have two separate flush functions, we can choose to flush
everythign to the OS, but don't enforce that it's physically written to the
disk.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
qcow2 has a writeback metadata cache, so flushing a qcow2 image actually
consists of writing back that cache to the protocol and only then flushes the
protocol in order to get everything stable on disk.
This introduces a separate bdrv_co_flush_to_os to reflect the split.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
There are two different types of flush that you can do: Flushing one level up
to the OS (i.e. writing data to the host page cache) or flushing it all the way
down to the disk. The existing functions flush to the disk, reflect this in the
function name.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Recent versions of udev always keep the tray locked so that the kernel
can observe "eject request" events (aka tray button presses) even on
discs that aren't mounted. Add support for these events in the ATAPI
and SCSI cd drive device models.
To let management cope with the behavior of udev, an event should also
be added for "tray opened/closed". This way, after issuing an "eject"
command, management can poll until the guests actually reacts to the
command. They can then issue the "change" command after the tray has been
opened, or try with "eject -f" after a (configurable?) timeout. However,
with this patch and the corresponding support in the device models,
at least it is possible to do a manual two-step eject+change sequence.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Several BlockDriverState fields are not being reinitialized across
bdrv_close()/bdrv_open(). Make sure they are reset to their default
values.
Signed-off-by: Stefan Hajnoczi <stefanha@linux.vnet.ibm.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Several block drivers set bs->read_only in .bdrv_open() but
block.c:bdrv_open_common() clobbers its value. Additionally, QED uses
bdrv_is_read_only() in .bdrv_open() to decide whether to perform
consistency checks.
The correct ordering is to initialize bs->read_only from the open flags
before calling .bdrv_open(). This way block drivers can override it if
necessary and can use bdrv_is_read_only() in .bdrv_open().
Signed-off-by: Stefan Hajnoczi <stefanha@linux.vnet.ibm.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
tmp_filename was used outside the block it was defined in, i.e. after it went
out of scope. Move its declaration to the top level.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Previous commits dropped most qobjects usage from qemu modules
(now they are a low level interface used by the QAPI). However,
some modules still include the qemu-objects.h header file.
This commit drops qemu-objects.h from some of those modules
and includes qjson.h instead, which is what they actually need.
Signed-off-by: Luiz Capitulino <lcapitulino@redhat.com>
The biggest change is to rename its prefix from BDRV_IOS to
BLOCK_DEVICE_IO_STATUS.
Next commit will convert the query-block command to the QAPI
and that's how the enumeration is going to be generated.
Signed-off-by: Luiz Capitulino <lcapitulino@redhat.com>
A future commit will convert bdrv_info() to the QAPI and it won't
provide IOS_INVAL.
Luckily all we have to do is to add a new 'iostatus_enabled'
member to BlockDriverState and use it instead.
Signed-off-by: Luiz Capitulino <lcapitulino@redhat.com>
Since coroutine operation is now mandatory, convert both bdrv_discard
implementations to coroutines. For qcow2, this means taking the lock
around the operation. raw-posix remains synchronous.
The bdrv_discard callback is then unused and can be eliminated.
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Since coroutine operation is now mandatory, convert all bdrv_flush
implementations to coroutines. For qcow2, this means taking the lock.
Other implementations are simpler and just forward bdrv_flush to the
underlying protocol, so they can avoid the lock.
The bdrv_flush callback is then unused and can be eliminated.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
This similarly adds support for coroutine and asynchronous discard.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Add coroutine support for flush and apply the same emulation that
we already do for read/write. bdrv_aio_flush is simplified to always
go through a coroutine.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Commit cd74d83345e0e3b708330ab8c4cd9111bb82cda6 ("block: switch
bdrv_read()/bdrv_write() to coroutines") removed the bdrv_has_async_rw()
callers. This patch removes bdrv_has_async_rw() since it is no longer
used.
Signed-off-by: Stefan Hajnoczi <stefanha@linux.vnet.ibm.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
There is no need to emulate .bdrv_read()/.bdrv_write() since these
interfaces are only called if aio and coroutine interfaces are not
present. All valid BlockDrivers must implement either sync, aio, or
coroutine interfaces.
Signed-off-by: Stefan Hajnoczi <stefanha@linux.vnet.ibm.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Block drivers that implement coroutine functions used to get sync and
aio wrappers. This is no longer necessary since all request processing
now happens in a coroutine. If a block driver implements the coroutine
interface then none of the other interfaces will be invoked.
Signed-off-by: Stefan Hajnoczi <stefanha@linux.vnet.ibm.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
More sync, aio, and coroutine unification. Make bdrv_aio_writev() go
through coroutine request processing.
Remove the dirty block callback mechanism which was needed only for aio
processing and can be done more naturally in coroutine context.
Signed-off-by: Stefan Hajnoczi <stefanha@linux.vnet.ibm.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
The aio write operation marks blocks dirty when the write operation
completes. The coroutine write operation marks blocks dirty before
issuing the write operation.
It seems safest to mark the block dirty when the operation completes so
that anything tracking dirty blocks will not act before the change has
been made to the image file.
Make the coroutine write operation dirty blocks on write completion.
Signed-off-by: Stefan Hajnoczi <stefanha@linux.vnet.ibm.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
More sync, aio, and coroutine unification. Make bdrv_aio_readv() go
through coroutine request processing.
Signed-off-by: Stefan Hajnoczi <stefanha@linux.vnet.ibm.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
The bdrv_read()/bdrv_write() functions call .bdrv_read()/.bdrv_write().
They should go through bdrv_co_do_readv() and bdrv_co_do_writev()
instead in order to unify request processing code across sync, aio, and
coroutine interfaces. This is also an important step towards removing
BlockDriverState .bdrv_read()/.bdrv_write() in the future.
Signed-off-by: Stefan Hajnoczi <stefanha@linux.vnet.ibm.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
The public interface for I/O in coroutine context is bdrv_co_readv() and
bdrv_co_writev(). Split out the request processing code into
bdrv_co_do_readv() and bdrv_co_writev() so that it can be called
internally when we refactor all request processing to use coroutines.
Signed-off-by: Stefan Hajnoczi <stefanha@linux.vnet.ibm.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
The emulation functions which supply default BlockDriver .bdrv_*()
functions given another implemented .bdrv_*() function should not use
public bdrv_*() interfaces. This patch ensures they invoke .bdrv_*()
directly to avoid adding an extra layer of coroutine request processing
and possibly entering an infinite loop.
Signed-off-by: Stefan Hajnoczi <stefanha@linux.vnet.ibm.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
We will unify block layer request processing across sync, aio, and
coroutines and this means a .bdrv_co_*() emulation function should not
call back into the public interface. There's no need here, just call
.bdrv_aio_*() directly.
The gory details: bdrv_co_io_em() cannot call back into the public
bdrv_aio_*() interface since that will be handled using coroutines,
which causes us to call into bdrv_co_io_em() again in an infinite loop
:).
Signed-off-by: Stefan Hajnoczi <stefanha@linux.vnet.ibm.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Luiz Capitulino <lcapitulino@redhat.com>
Reviewed-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Contains the I/O status for the given device. The key is only present
if the device supports it and the VM is configured to stop on errors.
Please, check the documentation being added in this commit for more
information.
Signed-off-by: Luiz Capitulino <lcapitulino@redhat.com>
Reviewed-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
This commit adds support to the BlockDriverState type to keep track
of devices' I/O status.
There are three possible status: BDRV_IOS_OK (no error), BDRV_IOS_ENOSPC
(no space error) and BDRV_IOS_FAILED (any other error). The distinction
between no space and other errors is important because a management
application may want to watch for no space in order to extend the
space assigned to the VM and put it to run again.
Qemu devices supporting the I/O status feature have to enable it
explicitly by calling bdrv_iostatus_enable() _and_ have to be
configured to stop the VM on errors (ie. werror=stop|enospc or
rerror=stop).
In case of multiple errors being triggered in sequence only the first
one is stored. The I/O status is always reset to BDRV_IOS_OK when the
'cont' command is issued.
Next commits will add support to some devices and extend the
query-block/info block commands to return the I/O status information.
Signed-off-by: Luiz Capitulino <lcapitulino@redhat.com>
Reviewed-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
It is useful to know the BlockDriverState as well as the
sector_num/nb_sectors of an emulated .bdrv_co_*() request.
Signed-off-by: Stefan Hajnoczi <stefanha@linux.vnet.ibm.com>
bdrv_open_common() is a useful point to trace since it reveals the
filename and block driver for a given BlockDriverState.
Signed-off-by: Stefan Hajnoczi <stefanha@linux.vnet.ibm.com>
To let device models distinguish between eject and load.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
BlockDriverState member buffer_alignment is initially 512. The device
model may set them, with bdrv_set_buffer_alignment(). If the device
model gets detached (hot unplug), the device's alignment is left
behind. Only okay because device hot unplug automatically destroys
the BlockDriverState. But that's a questionable feature, best not to
rely on it.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Device models should be able to set it without an unclean include of
block_int.h.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Need to ask the device, so this requires new BlockDevOps member
is_tray_open().
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
It's a confused mess (see previous commit). No users remain.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
BlockDriverState member removable is a confused mess. It is true when
an ide-cd, scsi-cd or floppy qdev is attached, or when the
BlockDriverState was created with -drive if={floppy,sd} or -drive
if={ide,scsi,xen,none},media=cdrom ("created removable"), except when
an ide-hd, scsi-hd, scsi-generic or virtio-blk qdev is attached.
Three users remain:
1. eject_device(), via bdrv_is_removable() uses it to determine
whether a block device can eject media.
2. bdrv_info() is monitor command "info block". QMP documentation
says "true if the device is removable, false otherwise". From the
monitor user's point of view, the only sensible interpretation of
"is removable" is "can eject media with monitor commands eject and
change".
A block device can eject media unless a device is attached that
doesn't support it. Switch the two users over to new
bdrv_dev_has_removable_media() that returns exactly that.
3. bdrv_getlength() uses to suppress its length cache when media can
change (see commit 46a4e4e6). Media change is either monitor
command change (updates the length cache), monitor command eject
(doesn't update the length cache, easily fixable), or physical
media change (invalidates length cache, not so easily fixable).
I'm refraining from improving anything here, because this series is
long enough already. Instead, I simply switch it over to
bdrv_dev_has_removable_media() as well.
This changes the behavior of the length cache and of monitor commands
eject and change in two cases:
a. drive not created removable, no device attached
The commit makes the drive removable, and defeats the length cache.
Example: -drive if=none
b. drive created removable, but the attached drive is non-removable,
and doesn't call bdrv_set_removable(..., 0) (most devices don't)
The commit makes the drive non-removable, and enables the length
cache.
Example: -drive if=xen,media=cdrom -M xenpv
The other non-removable devices that don't call
bdrv_set_removable() can't currently use a drive created removable,
either because they aren't qdevified, or because they lack a drive
property. Won't stay that way.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Requires new BlockDevOps member is_medium_locked(). Implement for IDE
and SCSI CD-ROMs.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
The device model knows best when to accept the guest's eject command.
No need to detour through the block layer.
bdrv_eject() can't fail anymore. Make it void.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Commit 4be9762a changed bdrv_is_inserted() to fail when the tray is
open. Unfortunately, there are two different kinds of users, with
conflicting needs.
1. Device models using bdrv_eject(), currently ide-cd and scsi-cd.
They expect bdrv_is_inserted() to reflect the tray status. Commit
4be9762a makes them happy.
2. Code that wants to know whether a BlockDriverState has media, such
as find_image_format(), bdrv_flush_all(). Commit 4be9762a makes them
unhappy. In particular, it breaks flush on VM stop for media ejected
by the guest.
Revert the change to bdrv_is_inserted(). Check the tray status in the
device models instead.
Note on IDE: Since only ATAPI devices have a tray, and they don't
accept ATA commands since the recent commit "ide: Reject ATA commands
specific to drive kinds", checking in atapi.c suffices.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
savevm and loadvm silently ignore block devices with removable media,
such as floppies and SD cards. Rolling back a VM to a previous
checkpoint will *not* roll back writes to block devices with removable
media.
Moreover, bdrv_is_removable() is a confused mess, and wrong in at
least one case: it considers "-drive if=xen,media=cdrom -M xenpv"
removable. It'll be cleaned up later in this series.
Read-only block devices are also ignored, but that's okay.
Fix by ignoring only read-only block devices and empty block devices.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Change (!bdrv_is_removable(bs) || bdrv_is_inserted(bs)) to just
bdrv_is_inserted(). Rationale:
The value of bdrv_is_removable(bs) matters only when
bdrv_is_inserted(bs) is false.
bdrv_is_inserted(bs) is true when bs is open (bs->drv != NULL) and
not an empty host drive (CD-ROM or floppy).
Therefore, bdrv_is_removable(bs) matters only when:
1. bs is not open
old: may call bdrv_flush(bs), which does nothing
new: won't call
2. bs is an empty host drive
old: may call bdrv_flush(bs), which calls driver method
raw_flush(), which calls fdatasync() or equivalent, which
can't do anything useful while the drive is empty
new: won't call
Result is bs->drv && !bdrv_is_read_only(bs) && bdrv_is_inserted(bs).
bdrv_is_inserted(bs) implies bs->drv. Drop the redundant test.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Multiplexing callbacks complicates matters needlessly.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
