Re: [Qemu-devel] [PATCH 1/2] Fix Block Hotplug race with drive_del()

[Markus Armbruster]

Ryan Harper <address@hidden> writes:

> * Markus Armbruster <address@hidden> [2010-11-10 11:40]:
>> Ryan Harper <address@hidden> writes:
>> 
>> > * Markus Armbruster <address@hidden> [2010-11-10 06:48]:
>> >> One real question, and a couple of nits.
>> >> 
>> >> Ryan Harper <address@hidden> writes:
>> >> 
>> >> > Block hot unplug is racy since the guest is required to acknowlege the 
>> >> > ACPI
>> >> > unplug event; this may not happen synchronously with the device removal 
>> >> > command
>> >> 
>> >> Well, I wouldn't call unplug "racy".  It just takes an unpredictable
>> >> length of time, possibly forever.  To make a race, you need to throw in
>> >> a client assuming (incorrectly) that unplug is instantaneous, as
>> >> described in your next paragraph.
>> >> 
>> >> Moreover, all PCI unplug is that way, not just block.
>> >> 
>> >> > This series aims to close a gap where by mgmt applications that assume 
>> >> > the
>> >> > block resource has been removed without confirming that the guest has
>> >> > acknowledged the removal may re-assign the underlying device to a 
>> >> > second guest
>> >> > leading to data leakage.
>> >> 
>> >> Yes, the incorrect assumption is a problem.  But with that fixed (in the
>> >> management application), we run right into the next problem: there is no
>> >> way for the management application to reliably disconnect the guest from
>> >> a block device.  And that's the problem you're fixing.
>> >
>> > Yeah, that's the right way to word it; providing a method to forcibly
>> > disconnect the guest from the host device.
>> >> 
>> >> > This series introduces a new montor command to decouple asynchornous 
>> >> > device
>> >> 
>> >> Typos "montor" and "asynchornous".  You might want to use a spell
>> >> checker :)
>> >> 
>> >> Lines are a bit long.  Recommend wrap at column 70.
>> >> 
>> >> > removal from restricting guest access to a block device.  We do this by 
>> >> > creating
>> >> > a new monitor command drive_del which maps to a bdrv_unplug() command 
>> >> > which
>> >> > does a qemu_aio_flush; bdrv_flush() and bdrv_close().  Once complete, 
>> >> > subsequent
>> >> > IO is rejected from the device and the guest will get IO errors but 
>> >> > continue to
>> >> > function.  In addition to preventing further IO, we clean up state 
>> >> > pointers
>> >> > between host (BlockDriverState) and guest (DeviceInfo).
>> >> >
>> >> > A subsequent device removal command can be issued to remove the device, 
>> >> > to which
>> >> > the guest may or maynot respond, but as long as the unplugged bit is 
>> >> > set, no IO
>> >> 
>> >> "maynot" is not a word.
>> >> 
>> >> > will be sumbitted.
>> >> 
>> >> This suggests to drive_del before device_del, which makes the device
>> >> goes through a "broken device" state on its way to unplug.  If the guest
>> >> accesses the device in that state, it gets I/O errors.  Not nice.
>> >> 
>> >> Instead, I'd recommend device_del, wait for the device to go away,
>> >> drive_del on time out.  If the guest reacts to the ACPI unplug promptly,
>> >> it's never exposed to the "broken device" state.  Note: if the drive_del
>> >> fails because the device doesn't exist, we lost the race with the
>> >> automatic destruction, which is harmless.  Ignore that error.
>> >
>> > Honestly, other than describing what happens if you sever the connection
>> > when the guest isn't aware of it; I don't want to try to capture how the
>> > mgmt layer implements the removal.  
>> >
>> > One may want to force the disconnect before attempting to remove the
>> > device; or the other way around; that's really the mgmt layer's call.
>> 
>> Fair enough.
>> 
>> >> > Signed-off-by: Ryan Harper <address@hidden>
>> >> > ---
>> >> >  block.c         |    7 +++++++
>> >> >  block.h         |    1 +
>> >> >  blockdev.c      |   36 ++++++++++++++++++++++++++++++++++++
>> >> >  blockdev.h      |    1 +
>> >> >  hmp-commands.hx |   18 ++++++++++++++++++
>> >> >  5 files changed, 63 insertions(+), 0 deletions(-)
>> >> >
>> >> > diff --git a/block.c b/block.c
>> >> > index 6b505fb..c76a796 100644
>> >> > --- a/block.c
>> >> > +++ b/block.c
>> >> > @@ -1328,6 +1328,13 @@ void bdrv_set_removable(BlockDriverState *bs, 
>> >> > int removable)
>> >> >      }
>> >> >  }
>> >> >  
>> >> > +void bdrv_unplug(BlockDriverState *bs)
>> >> > +{
>> >> > +    qemu_aio_flush();
>> >> > +    bdrv_flush(bs);
>> >> > +    bdrv_close(bs);
>> >> > +}
>> >> > +
>> >> 
>> >> Unless we expect more users, I'd inline this into its only caller.
>> >> Matter of taste.
>> >
>> > Works for me.
>> >
>> >> 
>> >> >  int bdrv_is_removable(BlockDriverState *bs)
>> >> >  {
>> >> >      return bs->removable;
>> >> > diff --git a/block.h b/block.h
>> >> > index 78ecfac..581414c 100644
>> >> > --- a/block.h
>> >> > +++ b/block.h
>> >> > @@ -171,6 +171,7 @@ void bdrv_set_on_error(BlockDriverState *bs, 
>> >> > BlockErrorAction on_read_error,
>> >> >                         BlockErrorAction on_write_error);
>> >> >  BlockErrorAction bdrv_get_on_error(BlockDriverState *bs, int is_read);
>> >> >  void bdrv_set_removable(BlockDriverState *bs, int removable);
>> >> > +void bdrv_unplug(BlockDriverState *bs);
>> >> >  int bdrv_is_removable(BlockDriverState *bs);
>> >> >  int bdrv_is_read_only(BlockDriverState *bs);
>> >> >  int bdrv_is_sg(BlockDriverState *bs);
>> >> > diff --git a/blockdev.c b/blockdev.c
>> >> > index 6cb179a..ee8c2ec 100644
>> >> > --- a/blockdev.c
>> >> > +++ b/blockdev.c
>> >> > @@ -14,6 +14,8 @@
>> >> >  #include "qemu-option.h"
>> >> >  #include "qemu-config.h"
>> >> >  #include "sysemu.h"
>> >> > +#include "hw/qdev.h"
>> >> > +#include "block_int.h"
>> >> >  
>> >> >  static QTAILQ_HEAD(drivelist, DriveInfo) drives = 
>> >> > QTAILQ_HEAD_INITIALIZER(drives);
>> >> >  
>> >> > @@ -597,3 +599,37 @@ int do_change_block(Monitor *mon, const char 
>> >> > *device,
>> >> >      }
>> >> >      return monitor_read_bdrv_key_start(mon, bs, NULL, NULL);
>> >> >  }
>> >> > +
>> >> > +int do_drive_del(Monitor *mon, const QDict *qdict, QObject **ret_data)
>> >> > +{
>> >> > +    const char *id = qdict_get_str(qdict, "id");
>> >> > +    BlockDriverState *bs;
>> >> > +    Property *prop;
>> >> > +
>> >> > +    bs = bdrv_find(id);
>> >> > +    if (!bs) {
>> >> > +        qerror_report(QERR_DEVICE_NOT_FOUND, id);
>> >> > +        return -1;
>> >> > +    }
>> >> > +
>> >> > +    /* quiesce block driver; prevent further io */
>> >> > +    bdrv_unplug(bs);
>> >> > +
>> >> > +    /* clean up guest state from pointing to host resource by
>> >> > +     * finding and removing DeviceState "drive" property */
>> >> > +    for (prop = bs->peer->info->props; prop && prop->name; prop++) {
>> >> > +        if ((prop->info->type == PROP_TYPE_DRIVE) && 
>> >> > +            (*(BlockDriverState **)qdev_get_prop_ptr(bs->peer, prop) 
>> >> > == bs)) {
>> >> > +            if (prop->info->free) {
>> >> > +                prop->info->free(bs->peer, prop);
>> >> > +            }
>> 
>> Your use of prop->info->free() in this context is wrong.  More below.
>> 
>> >> 
>> >> Does this null the drive property?  I doubt it.  Quick check in the
>> >> debugger?
>> >> 
>> >> The free callbacks generally don't zap the properties, because they run
>> >> from qdev_free().
>> >
>> > To be honest; I didn't see anything that looked like "remove this
>> > property" in the qdev api.  Any pointers?
>> 
>> The closest we have is indeed the Property method free(), but that's not
>> quite right.  It's really only for use by qdev_free().
>> 
>> > should I be calling qdev_free() on the dev?
>> 
>> No, because then the whole device is gone, not just the property :)
>> 
>> >                                              I don't quite understand
>> > the distinction between the info list of properties and the device
>> > itself, nor specifically what we need to remove in the drive_del()
>> > operation versus the device_del() portion.
>> 
>> device_del / qdev_free() destroy a qdev, such as a "virtio-blk-pci"
>> device (C type VirtIOPCIProxy).
>> 
>> drive_del destroys something else, namely the block device host part
>> (BlockDriverState + DeviceInfo).  Obviously, it needs to zap all
>> pointers to the host part along with it.  Specifically, it needs to zap
>> the device's pointer to it.
>> 
>> Example: if a "virtio-blk-pci" device is using drive "foo", then
>> "drive_del foo" needs to zap its member block.bs.
>> 
>> Complication: we don't (want to) know what kind of device exactly is
>> using the drive.  But we do know that a drive property must be
>> describing it.
>> 
>> So we search the properties (for (prop...)) for a drive property
>> (prop->info->type == PROP_TYPE_DRIVE) that points to this drive (... ==
>> bs).
>> 
>> Result:
>> 
>>     BlockDriverState *bs;
>>     Property *prop;
>>     BlockDriverState **ptr;
>> [...]
>>     for (prop = bs->peer->info->props; prop && prop->name; prop++) {
>>         if ((prop->info->type == PROP_TYPE_DRIVE)) {
>>             ptr = qdev_get_prop_ptr(dev, prop);
>>             if (*ptr == bs) {
>>                 bdrv_detach(bs, bs->peer);
>
> Invoking the free method on the drive property does do detach:
>
> free_drive
> {
>     BlockDriverState **ptr = qdev_get_prop_ptr(dev, prop);
>
>     if (*ptr) {
>         bdrv_detach(*ptr, dev);
>         blockdev_auto_del(*ptr);
>     }
> }
>
> and the bdrv_delete()
>
> takes out the bs pointer.

Which pointer?  Which bdrv_delete()?

>> Only then are we ready to destroy the host part:
>> 
>>     drive_uninit(drive_get_by_blockdev(bs));
>
> And if auto-deletion it set, then it handles the drive_uninit().  Do you think
> we should explicitly invoke drive_uninit() ?

Actually, blockdev_auto_del() deletes the block device only if DriveInfo
has auto_del set.  Why is that?  Quote blockdev.c:

/*
 * We automatically delete the drive when a device using it gets
 * unplugged.  Questionable feature, but we can't just drop it.
 * Device models call blockdev_mark_auto_del() to schedule the
 * automatic deletion, and generic qdev code calls blockdev_auto_del()
 * when deletion is actually safe.
 */

Thus, you need to blockdev_mark_auto_del() before blockdev_auto_del().

However, my blockdev_add work-in-progress changes these two functions to
*only* delete block devices created the old way (-drive, drive_add).
You want them deleted regardless of how they were created.  That's why I
asked you to use drive_uninit() directly.

You could argue that Property method free() *should* work here.  Fair
point.  If you want to clean that up, you're quite welcome.  But I don't
want to burden your fix with that, so feel free to add a suitable
comment instead.