mbox series

[v5,0/5] block: add a sequence number to disks

Message ID 20210712230530.29323-1-mcroce@linux.microsoft.com (mailing list archive)
Headers show
Series block: add a sequence number to disks | expand

Message

Matteo Croce July 12, 2021, 11:05 p.m. UTC
From: Matteo Croce <mcroce@microsoft.com>

Associating uevents with block devices in userspace is difficult and racy:
the uevent netlink socket is lossy, and on slow and overloaded systems has
a very high latency. Block devices do not have exclusive owners in
userspace, any process can set one up (e.g. loop devices). Moreover, device
names can be reused (e.g. loop0 can be reused again and again). A userspace
process setting up a block device and watching for its events cannot thus
reliably tell whether an event relates to the device it just set up or
another earlier instance with the same name.

Being able to set a UUID on a loop device would solve the race conditions.
But it does not allow to derive orderings from uevents: if you see a uevent
with a UUID that does not match the device you are waiting for, you cannot
tell whether it's because the right uevent has not arrived yet, or it was
already sent and you missed it. So you cannot tell whether you should wait
for it or not.

Being able to set devices up in a namespace would solve the race conditions
too, but it can work only if being namespaced is feasible in the first
place. Many userspace processes need to set devices up for the root
namespace, so this solution cannot always work.

Changing the loop devices naming implementation to always use
monotonically increasing device numbers, instead of reusing the lowest
free number, would also solve the problem, but it would be very disruptive
to userspace and likely break many existing use cases. It would also be
quite awkward to use on long-running machines, as the loop device name
would quickly grow to many-digits length.

Furthermore, this problem does not affect only loop devices - partition
probing is asynchronous and very slow on busy systems. It is very easy to
enter races when using LO_FLAGS_PARTSCAN and watching for the partitions to
show up, as it can take a long time for the uevents to be delivered after
setting them up.

Associating a unique, monotonically increasing sequential number to the
lifetime of each block device, which can be retrieved with an ioctl
immediately upon setting it up, allows to solve the race conditions with
uevents, and also allows userspace processes to know whether they should
wait for the uevent they need or if it was dropped and thus they should
move on.

This does not benefit only loop devices and block devices with multiple
partitions, but for example also removable media such as USB sticks or
cdroms/dvdroms/etc.

The first patch is the core one, the 2..4 expose the information in
different ways, and the last one makes the loop device generate a media
changed event upon attach, detach or reconfigure, so the sequence number
is increased.

If merged, this feature will immediately used by the userspace:
https://github.com/systemd/systemd/issues/17469#issuecomment-762919781

v4 -> v5:
- introduce a helper to raise media changed events
- use the new helper in loop instead of the full event code
- unexport inc_diskseq() which is only used by the block code now
- rebase on top of 5.14-rc1

v3 -> v4:
- rebased on top of 5.13
- hook the seqnum increase into the media change event
- make the loop device raise media change events
- merge 1/6 and 5/6
- move the uevent part of 1/6 into a separate one
- drop the now unneeded sysfs refactor
- change 'diskseq' to a global static variable
- add more comments
- refactor commit messages

v2 -> v3:
- rebased on top of 5.13-rc7
- resend because it appeared archived on patchwork

v1 -> v2:
- increase seqnum on media change
- increase on loop detach

Matteo Croce (6):
  block: add disk sequence number
  block: export the diskseq in uevents
  block: add ioctl to read the disk sequence number
  block: export diskseq in sysfs
  block: add a helper to raise a media changed event
  loop: raise media_change event

 Documentation/ABI/testing/sysfs-block | 12 ++++++
 block/disk-events.c                   | 62 +++++++++++++++++++++------
 block/genhd.c                         | 43 +++++++++++++++++++
 block/ioctl.c                         |  2 +
 drivers/block/loop.c                  |  5 +++
 include/linux/genhd.h                 |  3 ++
 include/uapi/linux/fs.h               |  1 +
 7 files changed, 114 insertions(+), 14 deletions(-)

Comments

Luca Boccassi July 20, 2021, 5:27 p.m. UTC | #1
On Tue, 2021-07-13 at 01:05 +0200, Matteo Croce wrote:
> From: Matteo Croce <mcroce@microsoft.com>
> 
> Associating uevents with block devices in userspace is difficult and racy:
> the uevent netlink socket is lossy, and on slow and overloaded systems has
> a very high latency. Block devices do not have exclusive owners in
> userspace, any process can set one up (e.g. loop devices). Moreover, device
> names can be reused (e.g. loop0 can be reused again and again). A userspace
> process setting up a block device and watching for its events cannot thus
> reliably tell whether an event relates to the device it just set up or
> another earlier instance with the same name.
> 
> Being able to set a UUID on a loop device would solve the race conditions.
> But it does not allow to derive orderings from uevents: if you see a uevent
> with a UUID that does not match the device you are waiting for, you cannot
> tell whether it's because the right uevent has not arrived yet, or it was
> already sent and you missed it. So you cannot tell whether you should wait
> for it or not.
> 
> Being able to set devices up in a namespace would solve the race conditions
> too, but it can work only if being namespaced is feasible in the first
> place. Many userspace processes need to set devices up for the root
> namespace, so this solution cannot always work.
> 
> Changing the loop devices naming implementation to always use
> monotonically increasing device numbers, instead of reusing the lowest
> free number, would also solve the problem, but it would be very disruptive
> to userspace and likely break many existing use cases. It would also be
> quite awkward to use on long-running machines, as the loop device name
> would quickly grow to many-digits length.
> 
> Furthermore, this problem does not affect only loop devices - partition
> probing is asynchronous and very slow on busy systems. It is very easy to
> enter races when using LO_FLAGS_PARTSCAN and watching for the partitions to
> show up, as it can take a long time for the uevents to be delivered after
> setting them up.
> 
> Associating a unique, monotonically increasing sequential number to the
> lifetime of each block device, which can be retrieved with an ioctl
> immediately upon setting it up, allows to solve the race conditions with
> uevents, and also allows userspace processes to know whether they should
> wait for the uevent they need or if it was dropped and thus they should
> move on.
> 
> This does not benefit only loop devices and block devices with multiple
> partitions, but for example also removable media such as USB sticks or
> cdroms/dvdroms/etc.
> 
> The first patch is the core one, the 2..4 expose the information in
> different ways, and the last one makes the loop device generate a media
> changed event upon attach, detach or reconfigure, so the sequence number
> is increased.
> 
> If merged, this feature will immediately used by the userspace:
> https://github.com/systemd/systemd/issues/17469#issuecomment-762919781
> 
> v4 -> v5:
> - introduce a helper to raise media changed events
> - use the new helper in loop instead of the full event code
> - unexport inc_diskseq() which is only used by the block code now
> - rebase on top of 5.14-rc1
> 
> v3 -> v4:
> - rebased on top of 5.13
> - hook the seqnum increase into the media change event
> - make the loop device raise media change events
> - merge 1/6 and 5/6
> - move the uevent part of 1/6 into a separate one
> - drop the now unneeded sysfs refactor
> - change 'diskseq' to a global static variable
> - add more comments
> - refactor commit messages
> 
> v2 -> v3:
> - rebased on top of 5.13-rc7
> - resend because it appeared archived on patchwork
> 
> v1 -> v2:
> - increase seqnum on media change
> - increase on loop detach
> 
> Matteo Croce (6):
>   block: add disk sequence number
>   block: export the diskseq in uevents
>   block: add ioctl to read the disk sequence number
>   block: export diskseq in sysfs
>   block: add a helper to raise a media changed event
>   loop: raise media_change event
> 
>  Documentation/ABI/testing/sysfs-block | 12 ++++++
>  block/disk-events.c                   | 62 +++++++++++++++++++++------
>  block/genhd.c                         | 43 +++++++++++++++++++
>  block/ioctl.c                         |  2 +
>  drivers/block/loop.c                  |  5 +++
>  include/linux/genhd.h                 |  3 ++
>  include/uapi/linux/fs.h               |  1 +
>  7 files changed, 114 insertions(+), 14 deletions(-)

For the series:

Tested-by: Luca Boccassi <bluca@debian.org>

I have implemented the basic systemd support for this (ioctl + uevent,
sysfs will be done later), and tested with this series on x86_64 and
Debian 11 userspace, everything seems to work great. Thanks Matteo!

Here's the implementation, in draft state until the kernel side is
merged:

https://github.com/systemd/systemd/pull/20257
Matteo Croce July 22, 2021, 11:41 a.m. UTC | #2
On Tue, Jul 20, 2021 at 7:27 PM Luca Boccassi <bluca@debian.org> wrote:
>
> On Tue, 2021-07-13 at 01:05 +0200, Matteo Croce wrote:
> > From: Matteo Croce <mcroce@microsoft.com>
> >
> > Associating uevents with block devices in userspace is difficult and racy:
> > the uevent netlink socket is lossy, and on slow and overloaded systems has
> > a very high latency. Block devices do not have exclusive owners in
> > userspace, any process can set one up (e.g. loop devices). Moreover, device
> > names can be reused (e.g. loop0 can be reused again and again). A userspace
> > process setting up a block device and watching for its events cannot thus
> > reliably tell whether an event relates to the device it just set up or
> > another earlier instance with the same name.
> >
> > Being able to set a UUID on a loop device would solve the race conditions.
> > But it does not allow to derive orderings from uevents: if you see a uevent
> > with a UUID that does not match the device you are waiting for, you cannot
> > tell whether it's because the right uevent has not arrived yet, or it was
> > already sent and you missed it. So you cannot tell whether you should wait
> > for it or not.
> >
> > Being able to set devices up in a namespace would solve the race conditions
> > too, but it can work only if being namespaced is feasible in the first
> > place. Many userspace processes need to set devices up for the root
> > namespace, so this solution cannot always work.
> >
> > Changing the loop devices naming implementation to always use
> > monotonically increasing device numbers, instead of reusing the lowest
> > free number, would also solve the problem, but it would be very disruptive
> > to userspace and likely break many existing use cases. It would also be
> > quite awkward to use on long-running machines, as the loop device name
> > would quickly grow to many-digits length.
> >
> > Furthermore, this problem does not affect only loop devices - partition
> > probing is asynchronous and very slow on busy systems. It is very easy to
> > enter races when using LO_FLAGS_PARTSCAN and watching for the partitions to
> > show up, as it can take a long time for the uevents to be delivered after
> > setting them up.
> >
> > Associating a unique, monotonically increasing sequential number to the
> > lifetime of each block device, which can be retrieved with an ioctl
> > immediately upon setting it up, allows to solve the race conditions with
> > uevents, and also allows userspace processes to know whether they should
> > wait for the uevent they need or if it was dropped and thus they should
> > move on.
> >
> > This does not benefit only loop devices and block devices with multiple
> > partitions, but for example also removable media such as USB sticks or
> > cdroms/dvdroms/etc.
> >
> > The first patch is the core one, the 2..4 expose the information in
> > different ways, and the last one makes the loop device generate a media
> > changed event upon attach, detach or reconfigure, so the sequence number
> > is increased.
> >
> > If merged, this feature will immediately used by the userspace:
> > https://github.com/systemd/systemd/issues/17469#issuecomment-762919781
> >
> > v4 -> v5:
> > - introduce a helper to raise media changed events
> > - use the new helper in loop instead of the full event code
> > - unexport inc_diskseq() which is only used by the block code now
> > - rebase on top of 5.14-rc1
> >
> > v3 -> v4:
> > - rebased on top of 5.13
> > - hook the seqnum increase into the media change event
> > - make the loop device raise media change events
> > - merge 1/6 and 5/6
> > - move the uevent part of 1/6 into a separate one
> > - drop the now unneeded sysfs refactor
> > - change 'diskseq' to a global static variable
> > - add more comments
> > - refactor commit messages
> >
> > v2 -> v3:
> > - rebased on top of 5.13-rc7
> > - resend because it appeared archived on patchwork
> >
> > v1 -> v2:
> > - increase seqnum on media change
> > - increase on loop detach
> >
> > Matteo Croce (6):
> >   block: add disk sequence number
> >   block: export the diskseq in uevents
> >   block: add ioctl to read the disk sequence number
> >   block: export diskseq in sysfs
> >   block: add a helper to raise a media changed event
> >   loop: raise media_change event
> >
> >  Documentation/ABI/testing/sysfs-block | 12 ++++++
> >  block/disk-events.c                   | 62 +++++++++++++++++++++------
> >  block/genhd.c                         | 43 +++++++++++++++++++
> >  block/ioctl.c                         |  2 +
> >  drivers/block/loop.c                  |  5 +++
> >  include/linux/genhd.h                 |  3 ++
> >  include/uapi/linux/fs.h               |  1 +
> >  7 files changed, 114 insertions(+), 14 deletions(-)
>
> For the series:
>
> Tested-by: Luca Boccassi <bluca@debian.org>
>
> I have implemented the basic systemd support for this (ioctl + uevent,
> sysfs will be done later), and tested with this series on x86_64 and
> Debian 11 userspace, everything seems to work great. Thanks Matteo!
>
> Here's the implementation, in draft state until the kernel side is
> merged:
>
> https://github.com/systemd/systemd/pull/20257
>

Hi Jens,

Given that the whole series has been acked and tested, and the
userspace has a draft implementation for it, is there anything else we
can do here?

Regards,
Lennart Poettering July 28, 2021, 7:01 p.m. UTC | #3
On Di, 20.07.21 18:27, Luca Boccassi (bluca@debian.org) wrote:

> Here's the implementation, in draft state until the kernel side is
> merged:
>
> https://github.com/systemd/systemd/pull/20257

I reviewed this systemd PR now. Looks excellent. See my comments on the PR.

Jens, anything we can do to get your blessing on the kernel patch set
and get this landed?

Thank you,

Lennart
Jens Axboe July 28, 2021, 7:22 p.m. UTC | #4
On 7/12/21 5:05 PM, Matteo Croce wrote:
> From: Matteo Croce <mcroce@microsoft.com>
> 
> Associating uevents with block devices in userspace is difficult and racy:
> the uevent netlink socket is lossy, and on slow and overloaded systems has
> a very high latency. Block devices do not have exclusive owners in
> userspace, any process can set one up (e.g. loop devices). Moreover, device
> names can be reused (e.g. loop0 can be reused again and again). A userspace
> process setting up a block device and watching for its events cannot thus
> reliably tell whether an event relates to the device it just set up or
> another earlier instance with the same name.
> 
> Being able to set a UUID on a loop device would solve the race conditions.
> But it does not allow to derive orderings from uevents: if you see a uevent
> with a UUID that does not match the device you are waiting for, you cannot
> tell whether it's because the right uevent has not arrived yet, or it was
> already sent and you missed it. So you cannot tell whether you should wait
> for it or not.
> 
> Being able to set devices up in a namespace would solve the race conditions
> too, but it can work only if being namespaced is feasible in the first
> place. Many userspace processes need to set devices up for the root
> namespace, so this solution cannot always work.
> 
> Changing the loop devices naming implementation to always use
> monotonically increasing device numbers, instead of reusing the lowest
> free number, would also solve the problem, but it would be very disruptive
> to userspace and likely break many existing use cases. It would also be
> quite awkward to use on long-running machines, as the loop device name
> would quickly grow to many-digits length.
> 
> Furthermore, this problem does not affect only loop devices - partition
> probing is asynchronous and very slow on busy systems. It is very easy to
> enter races when using LO_FLAGS_PARTSCAN and watching for the partitions to
> show up, as it can take a long time for the uevents to be delivered after
> setting them up.
> 
> Associating a unique, monotonically increasing sequential number to the
> lifetime of each block device, which can be retrieved with an ioctl
> immediately upon setting it up, allows to solve the race conditions with
> uevents, and also allows userspace processes to know whether they should
> wait for the uevent they need or if it was dropped and thus they should
> move on.
> 
> This does not benefit only loop devices and block devices with multiple
> partitions, but for example also removable media such as USB sticks or
> cdroms/dvdroms/etc.
> 
> The first patch is the core one, the 2..4 expose the information in
> different ways, and the last one makes the loop device generate a media
> changed event upon attach, detach or reconfigure, so the sequence number
> is increased.
> 
> If merged, this feature will immediately used by the userspace:
> https://github.com/systemd/systemd/issues/17469#issuecomment-762919781

Applied for 5.15, with #2 done manually since it didn't apply cleanly.